From dfccbf29f18d258602d88b597f6435a66f072f88 Mon Sep 17 00:00:00 2001
From: DK09876 <Dikshant.pradhan@vectorize.io>
Date: Mon, 15 Dec 2025 02:42:19 -0700
Subject: [PATCH] Added hindsight_liteLLM implementation (#17)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

* Added hindsight_liteLLM implementation

* Add instructions for entity vs bank id

* Add another line about entity

* Address PR review comments and enhance litellm integration

- Remove deprecated limit parameter from recall() and arecall() functions
  since Hindsight uses budget/max_tokens for result control
- Remove dead MODEL_MAX_OUTPUT_TOKENS dict and max_output_tokens property
  from LLMProvider (superseded by hardcoded max_completion_tokens)
- Add test-litellm-integration job to CI workflow
- Add reflect API support with use_reflect config option
- Add verbose mode debug info via get_last_injection_debug()
- Add entity_id support for multi-user memory isolation
- Add retain() and reflect() wrapper functions
- Update docstrings and examples

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* Make max_memories optional to allow unlimited memory injection

- Change max_memories default from 10 to None (no limit)
- When max_memories is None, all results from the API are used
- Fix recall result handling to properly detect list vs object return
- Update wrappers (OpenAI, Anthropic) with same optional behavior

This allows users to control memory limits via max_memory_tokens
and recall_budget without an artificial count limit.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* Remove entity_id from hindsight_litellm; add gpt-4o token cap

Multi-user support now uses separate bank_ids per user instead of
entity_id scoping (e.g., bank_id=f"user-{user_id}"). This simplifies
the API and aligns with the Hindsight architecture.

Also fixes max_completion_tokens error for gpt-4o models by capping
the value at 16384 (gpt-4o's limit) instead of sending the default
65000 which exceeds the model's supported maximum.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* Fix dark mode styling across Control Plane UI components

Improvements to ensure proper text visibility and contrast in both light
and dark modes:

- Add global CSS rules for datetime-local calendar picker icon visibility
  using filter: invert() for both light (0.5) and dark (1) modes
- Fix text colors in dialog components to use theme-aware foreground colors
- Update memory detail panel, document/chunk modals, and data views to use
  proper dark mode text classes (text-foreground, text-card-foreground)
- Fix form labels, headings, and content text in bank selector dialogs
- Update entities view and documents view table styling for dark mode
- Bump package versions to 0.1.4

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* Remove session_id feature and add How It Works section to README

- Remove session_id and session management (new_session, set_session,
  get_session) from config.py, callbacks.py, and __init__.py
- Session management was a client-only abstraction not backed by core API
- Add "How It Works" section to README with visual flow diagram
- Update README to remove session management documentation

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* Fix readme example

* Add dark mode again

---------

Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
---
 .github/workflows/test.yml                    |   29 +
 .../hindsight_api/engine/llm_wrapper.py       |    6 +-
 hindsight-api/hindsight_api/pg0.py            |   10 +-
 hindsight-control-plane/src/app/globals.css   |    9 +
 .../src/components/bank-profile-view.tsx      |    4 +-
 .../src/components/bank-selector.tsx          |   15 +-
 .../src/components/data-view.tsx              |   24 +-
 .../src/components/document-chunk-modal.tsx   |   22 +-
 .../src/components/documents-view.tsx         |   28 +-
 .../src/components/entities-view.tsx          |   16 +-
 .../src/components/memory-detail-panel.tsx    |   18 +-
 .../src/components/ui/dialog.tsx              |    4 +-
 hindsight-integrations/litellm/README.md      |  437 +++++++
 .../litellm/hindsight_litellm/__init__.py     |  817 ++++++++++++++
 .../litellm/hindsight_litellm/callbacks.py    |  640 +++++++++++
 .../litellm/hindsight_litellm/config.py       |  232 ++++
 .../litellm/hindsight_litellm/wrappers.py     | 1000 +++++++++++++++++
 hindsight-integrations/litellm/pyproject.toml |   59 +
 .../litellm/tests/__init__.py                 |    1 +
 .../litellm/tests/test_integration.py         |  471 ++++++++
 20 files changed, 3772 insertions(+), 70 deletions(-)
 create mode 100644 hindsight-integrations/litellm/README.md
 create mode 100644 hindsight-integrations/litellm/hindsight_litellm/__init__.py
 create mode 100644 hindsight-integrations/litellm/hindsight_litellm/callbacks.py
 create mode 100644 hindsight-integrations/litellm/hindsight_litellm/config.py
 create mode 100644 hindsight-integrations/litellm/hindsight_litellm/wrappers.py
 create mode 100644 hindsight-integrations/litellm/pyproject.toml
 create mode 100644 hindsight-integrations/litellm/tests/__init__.py
 create mode 100644 hindsight-integrations/litellm/tests/test_integration.py

diff --git a/.github/workflows/test.yml b/.github/workflows/test.yml
index 831b2626..6947fe3a 100644
--- a/.github/workflows/test.yml
+++ b/.github/workflows/test.yml
@@ -441,3 +441,32 @@ jobs:
       run: |
         echo "=== API Server Logs ==="
         cat /tmp/api-server.log || echo "No API server log found"
+
+  test-litellm-integration:
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Install uv
+      uses: astral-sh/setup-uv@v5
+      with:
+        enable-cache: true
+        prune-cache: false
+
+    - name: Set up Python
+      uses: actions/setup-python@v5
+      with:
+        python-version-file: ".python-version"
+
+    - name: Build litellm integration
+      working-directory: ./hindsight-integrations/litellm
+      run: uv build
+
+    - name: Install dependencies
+      working-directory: ./hindsight-integrations/litellm
+      run: uv sync --extra dev
+
+    - name: Run tests
+      working-directory: ./hindsight-integrations/litellm
+      run: uv run pytest tests -v
\ No newline at end of file
diff --git a/hindsight-api/hindsight_api/engine/llm_wrapper.py b/hindsight-api/hindsight_api/engine/llm_wrapper.py
index d3ff78ed..4d87c33f 100644
--- a/hindsight-api/hindsight_api/engine/llm_wrapper.py
+++ b/hindsight-api/hindsight_api/engine/llm_wrapper.py
@@ -175,9 +175,13 @@ class LLMProvider:
             is_reasoning_model = any(x in model_lower for x in ["gpt-5", "o1", "o3"])
 
             # For GPT-4 and GPT-4.1 models, cap max_completion_tokens to 32000
+            # For GPT-4o models, cap to 16384
             is_gpt4_model = any(x in model_lower for x in ["gpt-4.1", "gpt-4-"])
+            is_gpt4o_model = "gpt-4o" in model_lower
             if max_completion_tokens is not None:
-                if is_gpt4_model and max_completion_tokens > 32000:
+                if is_gpt4o_model and max_completion_tokens > 16384:
+                    max_completion_tokens = 16384
+                elif is_gpt4_model and max_completion_tokens > 32000:
                     max_completion_tokens = 32000
                 # For reasoning models, max_completion_tokens includes reasoning + output tokens
                 # Enforce minimum of 16000 to ensure enough space for both
diff --git a/hindsight-api/hindsight_api/pg0.py b/hindsight-api/hindsight_api/pg0.py
index ac642b52..1dc7ff6e 100644
--- a/hindsight-api/hindsight_api/pg0.py
+++ b/hindsight-api/hindsight_api/pg0.py
@@ -54,7 +54,9 @@ class EmbeddedPostgres:
                 loop = asyncio.get_event_loop()
                 info = await loop.run_in_executor(None, pg0.start)
                 logger.info(f"PostgreSQL started on port {self.port}")
-                return info.uri
+                # Construct URI manually since pg0-embedded may return None
+                uri = info.uri if info and info.uri else f"postgresql://{self.username}:{self.password}@localhost:{self.port}/{self.database}"
+                return uri
             except Exception as e:
                 last_error = str(e)
                 if attempt < max_retries:
@@ -89,9 +91,9 @@ class EmbeddedPostgres:
         pg0 = self._get_pg0()
         loop = asyncio.get_event_loop()
         info = await loop.run_in_executor(None, pg0.info)
-        if info is None or not info.running:
-            raise RuntimeError("PostgreSQL server is not running or URI not available")
-        return info.uri
+        # Construct URI manually since pg0-embedded may return None
+        uri = info.uri if info and info.uri else f"postgresql://{self.username}:{self.password}@localhost:{self.port}/{self.database}"
+        return uri
 
     async def is_running(self) -> bool:
         """Check if the PostgreSQL server is currently running."""
diff --git a/hindsight-control-plane/src/app/globals.css b/hindsight-control-plane/src/app/globals.css
index e41afbd5..b58199f1 100644
--- a/hindsight-control-plane/src/app/globals.css
+++ b/hindsight-control-plane/src/app/globals.css
@@ -180,4 +180,13 @@ code, pre {
   -webkit-background-clip: text;
   -webkit-text-fill-color: transparent;
   background-clip: text;
+}
+
+/* Fix datetime-local calendar icon visibility in both light and dark modes */
+input[type="datetime-local"]::-webkit-calendar-picker-indicator {
+  filter: invert(0.5);
+}
+
+.dark input[type="datetime-local"]::-webkit-calendar-picker-indicator {
+  filter: invert(1);
 }
\ No newline at end of file
diff --git a/hindsight-control-plane/src/components/bank-profile-view.tsx b/hindsight-control-plane/src/components/bank-profile-view.tsx
index 7a4896ce..aa823d86 100644
--- a/hindsight-control-plane/src/components/bank-profile-view.tsx
+++ b/hindsight-control-plane/src/components/bank-profile-view.tsx
@@ -239,7 +239,7 @@ export function BankProfileView() {
         <div className="flex gap-2">
           {editMode ? (
             <>
-              <Button onClick={handleCancel} variant="outline" disabled={saving}>
+              <Button onClick={handleCancel} variant="secondary" disabled={saving}>
                 Cancel
               </Button>
               <Button onClick={handleSave} disabled={saving}>
@@ -258,7 +258,7 @@ export function BankProfileView() {
             </>
           ) : (
             <>
-              <Button onClick={loadData} variant="outline" size="sm">
+              <Button onClick={loadData} variant="secondary" size="sm">
                 <RefreshCw className="w-4 h-4 mr-2" />
                 Refresh
               </Button>
diff --git a/hindsight-control-plane/src/components/bank-selector.tsx b/hindsight-control-plane/src/components/bank-selector.tsx
index 76b3aa60..8ff4573d 100644
--- a/hindsight-control-plane/src/components/bank-selector.tsx
+++ b/hindsight-control-plane/src/components/bank-selector.tsx
@@ -266,7 +266,7 @@ function BankSelectorInner() {
             </div>
             <DialogFooter>
               <Button
-                variant="outline"
+                variant="secondary"
                 onClick={() => {
                   setCreateDialogOpen(false);
                   setNewBankId('');
@@ -295,7 +295,7 @@ function BankSelectorInner() {
             </DialogHeader>
             <div className="py-4 space-y-4">
               <div>
-                <label className="font-bold block mb-1 text-sm">Content *</label>
+                <label className="font-bold block mb-1 text-sm text-foreground">Content *</label>
                 <Textarea
                   value={docContent}
                   onChange={(e) => setDocContent(e.target.value)}
@@ -306,7 +306,7 @@ function BankSelectorInner() {
               </div>
 
               <div>
-                <label className="font-bold block mb-1 text-sm">Context</label>
+                <label className="font-bold block mb-1 text-sm text-foreground">Context</label>
                 <Input
                   type="text"
                   value={docContext}
@@ -317,16 +317,17 @@ function BankSelectorInner() {
 
               <div className="grid grid-cols-2 gap-4">
                 <div>
-                  <label className="font-bold block mb-1 text-sm">Event Date</label>
+                  <label className="font-bold block mb-1 text-sm text-foreground">Event Date</label>
                   <Input
                     type="datetime-local"
                     value={docEventDate}
                     onChange={(e) => setDocEventDate(e.target.value)}
+                    className="text-foreground"
                   />
                 </div>
 
                 <div>
-                  <label className="font-bold block mb-1 text-sm">Document ID</label>
+                  <label className="font-bold block mb-1 text-sm text-foreground">Document ID</label>
                   <Input
                     type="text"
                     value={docDocumentId}
@@ -342,7 +343,7 @@ function BankSelectorInner() {
                   checked={docAsync}
                   onCheckedChange={(checked) => setDocAsync(checked as boolean)}
                 />
-                <label htmlFor="async-doc" className="text-sm cursor-pointer">
+                <label htmlFor="async-doc" className="text-sm cursor-pointer text-foreground">
                   Process in background (async)
                 </label>
               </div>
@@ -353,7 +354,7 @@ function BankSelectorInner() {
             </div>
             <DialogFooter>
               <Button
-                variant="outline"
+                variant="secondary"
                 onClick={() => {
                   setDocDialogOpen(false);
                   setDocContent('');
diff --git a/hindsight-control-plane/src/components/data-view.tsx b/hindsight-control-plane/src/components/data-view.tsx
index 503de63a..ca116fa8 100644
--- a/hindsight-control-plane/src/components/data-view.tsx
+++ b/hindsight-control-plane/src/components/data-view.tsx
@@ -480,7 +480,7 @@ export function DataView({ factType }: DataViewProps) {
                                       }`}
                                     >
                                       <TableCell className="py-2">
-                                        <div className="line-clamp-2 text-sm leading-snug">{row.text}</div>
+                                        <div className="line-clamp-2 text-sm leading-snug text-foreground">{row.text}</div>
                                         {row.context && (
                                           <div className="text-xs text-muted-foreground mt-0.5 truncate">{row.context}</div>
                                         )}
@@ -506,10 +506,10 @@ export function DataView({ factType }: DataViewProps) {
                                           <span className="text-xs text-muted-foreground">-</span>
                                         )}
                                       </TableCell>
-                                      <TableCell className="text-xs py-2">
+                                      <TableCell className="text-xs py-2 text-foreground">
                                         {occurredDisplay || <span className="text-muted-foreground">-</span>}
                                       </TableCell>
-                                      <TableCell className="text-xs py-2">
+                                      <TableCell className="text-xs py-2 text-foreground">
                                         {mentionedDisplay || <span className="text-muted-foreground">-</span>}
                                       </TableCell>
                                       <TableCell className="py-2">
@@ -519,7 +519,7 @@ export function DataView({ factType }: DataViewProps) {
                                             copyToClipboard(row.id);
                                           }}
                                           size="sm"
-                                          variant="ghost"
+                                          variant="secondary"
                                           className="h-6 w-6 p-0"
                                           title="Copy ID"
                                         >
@@ -799,7 +799,7 @@ function TimelineView({ data, filteredRows }: { data: any; filteredRows: any[] }
             {/* Zoom controls */}
             <div className="flex items-center border border-border rounded mr-2">
               <Button
-                variant="ghost"
+                variant="secondary"
                 size="sm"
                 onClick={zoomOut}
                 disabled={granularity === 'year'}
@@ -808,11 +808,11 @@ function TimelineView({ data, filteredRows }: { data: any; filteredRows: any[] }
               >
                 <ZoomOut className="h-3 w-3" />
               </Button>
-              <span className="text-[10px] px-2 min-w-[50px] text-center border-x border-border">
+              <span className="text-[10px] px-2 min-w-[50px] text-center border-x border-border text-foreground">
                 {granularityLabels[granularity]}
               </span>
               <Button
-                variant="ghost"
+                variant="secondary"
                 size="sm"
                 onClick={zoomIn}
                 disabled={granularity === 'day'}
@@ -826,7 +826,7 @@ function TimelineView({ data, filteredRows }: { data: any; filteredRows: any[] }
             {/* Navigation controls */}
             <div className="flex items-center border border-border rounded">
               <Button
-                variant="ghost"
+                variant="secondary"
                 size="sm"
                 onClick={() => scrollToGroup(0)}
                 disabled={timelineGroups.length <= 1}
@@ -836,7 +836,7 @@ function TimelineView({ data, filteredRows }: { data: any; filteredRows: any[] }
                 <ChevronsLeft className="h-3 w-3" />
               </Button>
               <Button
-                variant="ghost"
+                variant="secondary"
                 size="sm"
                 onClick={() => scrollToGroup(currentIndex - 1)}
                 disabled={currentIndex === 0}
@@ -845,11 +845,11 @@ function TimelineView({ data, filteredRows }: { data: any; filteredRows: any[] }
               >
                 <ChevronLeft className="h-3 w-3" />
               </Button>
-              <span className="text-[10px] px-2 min-w-[60px] text-center border-x border-border">
+              <span className="text-[10px] px-2 min-w-[60px] text-center border-x border-border text-foreground">
                 {currentIndex + 1} / {timelineGroups.length}
               </span>
               <Button
-                variant="ghost"
+                variant="secondary"
                 size="sm"
                 onClick={() => scrollToGroup(currentIndex + 1)}
                 disabled={currentIndex >= timelineGroups.length - 1}
@@ -859,7 +859,7 @@ function TimelineView({ data, filteredRows }: { data: any; filteredRows: any[] }
                 <ChevronRight className="h-3 w-3" />
               </Button>
               <Button
-                variant="ghost"
+                variant="secondary"
                 size="sm"
                 onClick={() => scrollToGroup(timelineGroups.length - 1)}
                 disabled={timelineGroups.length <= 1}
diff --git a/hindsight-control-plane/src/components/document-chunk-modal.tsx b/hindsight-control-plane/src/components/document-chunk-modal.tsx
index e27b786d..e2c8a127 100644
--- a/hindsight-control-plane/src/components/document-chunk-modal.tsx
+++ b/hindsight-control-plane/src/components/document-chunk-modal.tsx
@@ -94,7 +94,7 @@ export function DocumentChunkModal({ type, id, onClose }: DocumentChunkModalProp
                       <div className="text-xs font-bold text-muted-foreground uppercase mb-1">
                         Document ID
                       </div>
-                      <div className="text-sm font-mono break-all">{data.id}</div>
+                      <div className="text-sm font-mono break-all text-foreground">{data.id}</div>
                     </div>
                     {data.created_at && (
                       <div className="grid grid-cols-2 gap-3">
@@ -102,7 +102,7 @@ export function DocumentChunkModal({ type, id, onClose }: DocumentChunkModalProp
                           <div className="text-xs font-bold text-muted-foreground uppercase mb-1">
                             Created
                           </div>
-                          <div className="text-sm">
+                          <div className="text-sm text-foreground">
                             {new Date(data.created_at).toLocaleString()}
                           </div>
                         </div>
@@ -110,7 +110,7 @@ export function DocumentChunkModal({ type, id, onClose }: DocumentChunkModalProp
                           <div className="text-xs font-bold text-muted-foreground uppercase mb-1">
                             Memory Units
                           </div>
-                          <div className="text-sm">{data.memory_unit_count}</div>
+                          <div className="text-sm text-foreground">{data.memory_unit_count}</div>
                         </div>
                       </div>
                     )}
@@ -119,7 +119,7 @@ export function DocumentChunkModal({ type, id, onClose }: DocumentChunkModalProp
                         <div className="text-xs font-bold text-muted-foreground uppercase mb-1">
                           Text Length
                         </div>
-                        <div className="text-sm">
+                        <div className="text-sm text-foreground">
                           {data.original_text.length.toLocaleString()} characters
                         </div>
                       </div>
@@ -132,7 +132,7 @@ export function DocumentChunkModal({ type, id, onClose }: DocumentChunkModalProp
                         Original Text
                       </div>
                       <div className="p-4 bg-muted rounded-lg border border-border max-h-[300px] overflow-y-auto">
-                        <pre className="text-sm whitespace-pre-wrap font-mono">
+                        <pre className="text-sm whitespace-pre-wrap font-mono text-foreground">
                           {data.original_text}
                         </pre>
                       </div>
@@ -146,7 +146,7 @@ export function DocumentChunkModal({ type, id, onClose }: DocumentChunkModalProp
                       <div className="text-xs font-bold text-muted-foreground uppercase mb-1">
                         Chunk ID
                       </div>
-                      <div className="text-sm font-mono break-all">
+                      <div className="text-sm font-mono break-all text-foreground">
                         {data.chunk_id}
                       </div>
                     </div>
@@ -155,7 +155,7 @@ export function DocumentChunkModal({ type, id, onClose }: DocumentChunkModalProp
                         <div className="text-xs font-bold text-muted-foreground uppercase mb-1">
                           Document ID
                         </div>
-                        <div className="text-sm font-mono break-all">
+                        <div className="text-sm font-mono break-all text-foreground">
                           {data.document_id}
                         </div>
                       </div>
@@ -163,7 +163,7 @@ export function DocumentChunkModal({ type, id, onClose }: DocumentChunkModalProp
                         <div className="text-xs font-bold text-muted-foreground uppercase mb-1">
                           Chunk Index
                         </div>
-                        <div className="text-sm">{data.chunk_index}</div>
+                        <div className="text-sm text-foreground">{data.chunk_index}</div>
                       </div>
                     </div>
                     {data.created_at && (
@@ -171,7 +171,7 @@ export function DocumentChunkModal({ type, id, onClose }: DocumentChunkModalProp
                         <div className="text-xs font-bold text-muted-foreground uppercase mb-1">
                           Created
                         </div>
-                        <div className="text-sm">
+                        <div className="text-sm text-foreground">
                           {new Date(data.created_at).toLocaleString()}
                         </div>
                       </div>
@@ -181,7 +181,7 @@ export function DocumentChunkModal({ type, id, onClose }: DocumentChunkModalProp
                         <div className="text-xs font-bold text-muted-foreground uppercase mb-1">
                           Text Length
                         </div>
-                        <div className="text-sm">
+                        <div className="text-sm text-foreground">
                           {data.chunk_text.length.toLocaleString()} characters
                         </div>
                       </div>
@@ -194,7 +194,7 @@ export function DocumentChunkModal({ type, id, onClose }: DocumentChunkModalProp
                         Chunk Text
                       </div>
                       <div className="p-4 bg-muted rounded-lg border border-border max-h-[300px] overflow-y-auto">
-                        <pre className="text-sm whitespace-pre-wrap font-mono">
+                        <pre className="text-sm whitespace-pre-wrap font-mono text-foreground">
                           {data.chunk_text}
                         </pre>
                       </div>
diff --git a/hindsight-control-plane/src/components/documents-view.tsx b/hindsight-control-plane/src/components/documents-view.tsx
index ad431ea1..7cdfbb7c 100644
--- a/hindsight-control-plane/src/components/documents-view.tsx
+++ b/hindsight-control-plane/src/components/documents-view.tsx
@@ -122,17 +122,17 @@ export function DocumentsView() {
                         className={`cursor-pointer hover:bg-muted/50 ${selectedDocument?.id === doc.id ? 'bg-primary/10' : ''}`}
                         onClick={() => viewDocumentText(doc.id)}
                       >
-                        <TableCell title={doc.id}>
+                        <TableCell title={doc.id} className="text-card-foreground">
                           {doc.id.length > 30 ? doc.id.substring(0, 30) + '...' : doc.id}
                         </TableCell>
-                        <TableCell>
+                        <TableCell className="text-card-foreground">
                           {doc.created_at ? new Date(doc.created_at).toLocaleString() : 'N/A'}
                         </TableCell>
-                        <TableCell>
+                        <TableCell className="text-card-foreground">
                           {doc.retain_params?.context || '-'}
                         </TableCell>
-                        <TableCell>{doc.text_length?.toLocaleString()} chars</TableCell>
-                        <TableCell>{doc.memory_unit_count}</TableCell>
+                        <TableCell className="text-card-foreground">{doc.text_length?.toLocaleString()} chars</TableCell>
+                        <TableCell className="text-card-foreground">{doc.memory_unit_count}</TableCell>
                         <TableCell>
                           <Button
                             onClick={(e) => {
@@ -140,7 +140,7 @@ export function DocumentsView() {
                               viewDocumentText(doc.id);
                             }}
                             size="sm"
-                            variant={selectedDocument?.id === doc.id ? 'default' : 'outline'}
+                            variant={selectedDocument?.id === doc.id ? 'default' : 'secondary'}
                             title="View original text"
                           >
                             View Text
@@ -171,7 +171,7 @@ export function DocumentsView() {
                     <p className="text-sm text-muted-foreground mt-1">Original document text and metadata</p>
                   </div>
                   <Button
-                    variant="outline"
+                    variant="secondary"
                     size="sm"
                     onClick={() => setSelectedDocument(null)}
                     className="h-9 px-3 gap-2"
@@ -193,7 +193,7 @@ export function DocumentsView() {
                     {/* Document ID */}
                     <div className="p-4 bg-muted/50 rounded-lg">
                       <div className="text-xs font-bold text-muted-foreground uppercase mb-2">Document ID</div>
-                      <code className="text-sm font-mono break-all text-foreground">{selectedDocument.id}</code>
+                      <div className="text-sm font-mono break-all text-card-foreground">{selectedDocument.id}</div>
                     </div>
 
                     {/* Created & Memory Units */}
@@ -201,11 +201,11 @@ export function DocumentsView() {
                       <div className="grid grid-cols-2 gap-4">
                         <div className="p-4 bg-muted/50 rounded-lg">
                           <div className="text-xs font-bold text-muted-foreground uppercase mb-2">Created</div>
-                          <div className="text-sm font-medium text-foreground">{new Date(selectedDocument.created_at).toLocaleString()}</div>
+                          <div className="text-sm font-medium text-card-foreground">{new Date(selectedDocument.created_at).toLocaleString()}</div>
                         </div>
                         <div className="p-4 bg-muted/50 rounded-lg">
                           <div className="text-xs font-bold text-muted-foreground uppercase mb-2">Memory Units</div>
-                          <div className="text-sm font-medium text-foreground">{selectedDocument.memory_unit_count}</div>
+                          <div className="text-sm font-medium text-card-foreground">{selectedDocument.memory_unit_count}</div>
                         </div>
                       </div>
                     )}
@@ -214,7 +214,7 @@ export function DocumentsView() {
                     {selectedDocument.original_text && (
                       <div className="p-4 bg-muted/50 rounded-lg">
                         <div className="text-xs font-bold text-muted-foreground uppercase mb-2">Text Length</div>
-                        <div className="text-sm font-medium text-foreground">{selectedDocument.original_text.length.toLocaleString()} characters</div>
+                        <div className="text-sm font-medium text-card-foreground">{selectedDocument.original_text.length.toLocaleString()} characters</div>
                       </div>
                     )}
 
@@ -222,7 +222,7 @@ export function DocumentsView() {
                     {selectedDocument.retain_params && (
                       <div className="p-4 bg-muted/50 rounded-lg">
                         <div className="text-xs font-bold text-muted-foreground uppercase mb-2">Retain Parameters</div>
-                        <div className="text-sm space-y-2">
+                        <div className="text-sm space-y-2 text-card-foreground">
                           {selectedDocument.retain_params.context && (
                             <div><span className="font-semibold">Context:</span> {selectedDocument.retain_params.context}</div>
                           )}
@@ -232,7 +232,7 @@ export function DocumentsView() {
                           {selectedDocument.retain_params.metadata && (
                             <div className="mt-2">
                               <span className="font-semibold">Metadata:</span>
-                              <pre className="mt-1 text-xs bg-background p-2 rounded">{JSON.stringify(selectedDocument.retain_params.metadata, null, 2)}</pre>
+                              <pre className="mt-1 text-xs bg-background p-2 rounded text-card-foreground">{JSON.stringify(selectedDocument.retain_params.metadata, null, 2)}</pre>
                             </div>
                           )}
                         </div>
@@ -244,7 +244,7 @@ export function DocumentsView() {
                       <div>
                         <div className="text-xs font-bold text-muted-foreground uppercase mb-2">Original Text</div>
                         <div className="p-4 bg-muted/50 rounded-lg border border-border max-h-[400px] overflow-y-auto">
-                          <pre className="text-sm whitespace-pre-wrap font-mono leading-relaxed text-foreground">{selectedDocument.original_text}</pre>
+                          <pre className="text-sm whitespace-pre-wrap font-mono leading-relaxed text-card-foreground">{selectedDocument.original_text}</pre>
                         </div>
                       </div>
                     )}
diff --git a/hindsight-control-plane/src/components/entities-view.tsx b/hindsight-control-plane/src/components/entities-view.tsx
index 767cc976..bd130c36 100644
--- a/hindsight-control-plane/src/components/entities-view.tsx
+++ b/hindsight-control-plane/src/components/entities-view.tsx
@@ -126,10 +126,10 @@ export function EntitiesView() {
                       selectedEntity?.id === entity.id ? 'bg-primary/10' : ''
                     }`}
                   >
-                    <TableCell className="font-medium">{entity.canonical_name}</TableCell>
-                    <TableCell>{entity.mention_count}</TableCell>
-                    <TableCell>{formatDate(entity.first_seen)}</TableCell>
-                    <TableCell>{formatDate(entity.last_seen)}</TableCell>
+                    <TableCell className="font-medium text-card-foreground">{entity.canonical_name}</TableCell>
+                    <TableCell className="text-card-foreground">{entity.mention_count}</TableCell>
+                    <TableCell className="text-card-foreground">{formatDate(entity.first_seen)}</TableCell>
+                    <TableCell className="text-card-foreground">{formatDate(entity.last_seen)}</TableCell>
                   </TableRow>
                 ))}
               </TableBody>
@@ -154,7 +154,7 @@ export function EntitiesView() {
             {/* Header */}
             <div className="flex justify-between items-center mb-6 pb-4 border-b border-border">
               <div>
-                <h3 className="text-xl font-bold text-foreground">{selectedEntity.canonical_name}</h3>
+                <h3 className="text-xl font-bold text-card-foreground">{selectedEntity.canonical_name}</h3>
                 <p className="text-sm text-muted-foreground mt-1">Entity details</p>
               </div>
               <Button
@@ -172,11 +172,11 @@ export function EntitiesView() {
               <div className="grid grid-cols-2 gap-4">
                 <div className="p-4 bg-muted/50 rounded-lg">
                   <div className="text-xs font-bold text-muted-foreground uppercase mb-2">Mentions</div>
-                  <div className="text-lg font-semibold text-foreground">{selectedEntity.mention_count}</div>
+                  <div className="text-lg font-semibold text-card-foreground">{selectedEntity.mention_count}</div>
                 </div>
                 <div className="p-4 bg-muted/50 rounded-lg">
                   <div className="text-xs font-bold text-muted-foreground uppercase mb-2">First Seen</div>
-                  <div className="text-sm font-medium text-foreground">{formatDate(selectedEntity.first_seen)}</div>
+                  <div className="text-sm font-medium text-card-foreground">{formatDate(selectedEntity.first_seen)}</div>
                 </div>
               </div>
 
@@ -206,7 +206,7 @@ export function EntitiesView() {
                   <ul className="space-y-2">
                     {selectedEntity.observations.map((obs, idx) => (
                       <li key={idx} className="p-3 bg-muted/50 rounded-lg">
-                        <div className="text-sm text-foreground">{obs.text}</div>
+                        <div className="text-sm text-card-foreground">{obs.text}</div>
                         {obs.mentioned_at && (
                           <div className="text-xs text-muted-foreground mt-2">
                             {formatDate(obs.mentioned_at)}
diff --git a/hindsight-control-plane/src/components/memory-detail-panel.tsx b/hindsight-control-plane/src/components/memory-detail-panel.tsx
index 87abb7df..10e96452 100644
--- a/hindsight-control-plane/src/components/memory-detail-panel.tsx
+++ b/hindsight-control-plane/src/components/memory-detail-panel.tsx
@@ -67,7 +67,7 @@ export function MemoryDetailPanel({
               <p className="text-sm text-muted-foreground mt-1">Full memory content and metadata</p>
             </div>
             <Button
-              variant="ghost"
+              variant="secondary"
               size="sm"
               onClick={onClose}
               className="h-8 w-8 p-0"
@@ -80,14 +80,14 @@ export function MemoryDetailPanel({
             {/* Full Text */}
             <div>
               <div className="text-xs font-bold text-muted-foreground uppercase mb-2">Full Text</div>
-              <div className="text-sm whitespace-pre-wrap leading-relaxed">{memory.text}</div>
+              <div className="text-sm whitespace-pre-wrap leading-relaxed text-foreground">{memory.text}</div>
             </div>
 
             {/* Context */}
             {memory.context && (
               <div className="p-4 bg-muted/50 rounded-lg">
                 <div className="text-xs font-bold text-muted-foreground uppercase mb-2">Context</div>
-                <div className="text-sm">{memory.context}</div>
+                <div className="text-sm text-foreground">{memory.context}</div>
               </div>
             )}
 
@@ -95,7 +95,7 @@ export function MemoryDetailPanel({
             <div className="grid grid-cols-2 gap-4">
               <div className="p-4 bg-muted/50 rounded-lg">
                 <div className="text-xs font-bold text-muted-foreground uppercase mb-2">Occurred</div>
-                <div className="text-sm font-medium">
+                <div className="text-sm font-medium text-foreground">
                   {memory.occurred_start
                     ? new Date(memory.occurred_start).toLocaleString()
                     : 'N/A'}
@@ -103,7 +103,7 @@ export function MemoryDetailPanel({
               </div>
               <div className="p-4 bg-muted/50 rounded-lg">
                 <div className="text-xs font-bold text-muted-foreground uppercase mb-2">Mentioned</div>
-                <div className="text-sm font-medium">
+                <div className="text-sm font-medium text-foreground">
                   {memory.mentioned_at
                     ? new Date(memory.mentioned_at).toLocaleString()
                     : 'N/A'}
@@ -159,7 +159,7 @@ export function MemoryDetailPanel({
                 {memory.document_id && (
                   <Button
                     onClick={() => openDocumentModal(memory.document_id)}
-                    variant="outline"
+                    variant="secondary"
                     className="flex-1"
                   >
                     View Document
@@ -168,7 +168,7 @@ export function MemoryDetailPanel({
                 {memory.chunk_id && (
                   <Button
                     onClick={() => openChunkModal(memory.chunk_id)}
-                    variant="outline"
+                    variant="secondary"
                     className="flex-1"
                   >
                     View Chunk
@@ -300,7 +300,7 @@ export function MemoryDetailPanel({
                 <Button
                   onClick={() => openDocumentModal(memory.document_id)}
                   size="sm"
-                  variant="outline"
+                  variant="secondary"
                   className={`flex-1 ${compact ? 'h-7 text-xs' : ''}`}
                 >
                   View Document
@@ -310,7 +310,7 @@ export function MemoryDetailPanel({
                 <Button
                   onClick={() => openChunkModal(memory.chunk_id)}
                   size="sm"
-                  variant="outline"
+                  variant="secondary"
                   className={`flex-1 ${compact ? 'h-7 text-xs' : ''}`}
                 >
                   View Chunk
diff --git a/hindsight-control-plane/src/components/ui/dialog.tsx b/hindsight-control-plane/src/components/ui/dialog.tsx
index f38593bd..45c6f1bd 100644
--- a/hindsight-control-plane/src/components/ui/dialog.tsx
+++ b/hindsight-control-plane/src/components/ui/dialog.tsx
@@ -44,7 +44,7 @@ const DialogContent = React.forwardRef<
       {...props}
     >
       {children}
-      <DialogPrimitive.Close className="absolute right-4 top-4 rounded-sm opacity-70 ring-offset-background transition-opacity hover:opacity-100 focus:outline-none focus:ring-2 focus:ring-ring focus:ring-offset-2 disabled:pointer-events-none data-[state=open]:bg-accent data-[state=open]:text-muted-foreground">
+      <DialogPrimitive.Close className="absolute right-4 top-4 rounded-sm ring-offset-background transition-opacity hover:opacity-80 focus:outline-none focus:ring-2 focus:ring-ring focus:ring-offset-2 disabled:pointer-events-none data-[state=open]:bg-accent data-[state=open]:text-muted-foreground text-foreground">
         <X className="h-4 w-4" />
         <span className="sr-only">Close</span>
       </DialogPrimitive.Close>
@@ -88,7 +88,7 @@ const DialogTitle = React.forwardRef<
   <DialogPrimitive.Title
     ref={ref}
     className={cn(
-      "text-lg font-semibold leading-none tracking-tight",
+      "text-lg font-semibold leading-none tracking-tight text-foreground",
       className
     )}
     {...props}
diff --git a/hindsight-integrations/litellm/README.md b/hindsight-integrations/litellm/README.md
new file mode 100644
index 00000000..d484cba1
--- /dev/null
+++ b/hindsight-integrations/litellm/README.md
@@ -0,0 +1,437 @@
+# hindsight-litellm
+
+Universal LLM memory integration via LiteLLM. Add persistent memory to any LLM application with just a few lines of code.
+
+## Features
+
+- **Universal LLM Support** - Works with 100+ LLM providers via LiteLLM (OpenAI, Anthropic, Groq, Azure, AWS Bedrock, Google Vertex AI, and more)
+- **Simple Integration** - Just configure, enable, and use `hindsight_litellm.completion()`
+- **Automatic Memory Injection** - Relevant memories are injected into prompts before LLM calls
+- **Automatic Conversation Storage** - Conversations are stored to Hindsight for future recall
+- **Two Memory Modes** - Choose between `reflect` (synthesized context) or `recall` (raw memory retrieval)
+- **Direct Memory APIs** - Query, synthesize, and store memories manually
+- **Native Client Wrappers** - Alternative wrappers for OpenAI and Anthropic SDKs
+- **Debug Mode** - Inspect exactly what memories are being injected
+
+## Installation
+
+```bash
+pip install hindsight-litellm
+```
+
+## Quick Start
+
+```python
+import hindsight_litellm
+
+# Configure and enable memory integration
+hindsight_litellm.configure(
+    hindsight_api_url="http://localhost:8888",
+    bank_id="my-agent",
+)
+hindsight_litellm.enable()
+
+# Use the convenience wrapper - memory is automatically injected and stored
+response = hindsight_litellm.completion(
+    model="gpt-4o-mini",
+    messages=[{"role": "user", "content": "What did we discuss about AI?"}]
+)
+```
+
+## How It Works
+
+Here's what happens under the hood when you call `completion()`:
+
+```
+┌─────────────────────────────────────────────────────────────────────────────┐
+│  1. YOUR CODE                                                               │
+│  ───────────────────────────────────────────────────────────────────────── │
+│  response = hindsight_litellm.completion(                                   │
+│      model="gpt-4o-mini",                                                   │
+│      messages=[{"role": "user", "content": "Help me with my Python project"}]│
+│  )                                                                          │
+└─────────────────────────────────────────────────────────────────────────────┘
+                                      │
+                                      ▼
+┌─────────────────────────────────────────────────────────────────────────────┐
+│  2. MEMORY RETRIEVAL (before LLM call)                                      │
+│  ───────────────────────────────────────────────────────────────────────── │
+│  # hindsight_litellm queries Hindsight for relevant memories                │
+│                                                                             │
+│  # If use_reflect=False (default) - raw memories:                           │
+│  memories = hindsight.recall(query="Help me with my Python project")        │
+│  # Returns: ["User prefers pytest", "User is building a FastAPI app", ...]  │
+│                                                                             │
+│  # If use_reflect=True - synthesized context:                               │
+│  context = hindsight.reflect(query="Help me with my Python project")        │
+│  # Returns: "The user is an experienced Python developer working on..."     │
+└─────────────────────────────────────────────────────────────────────────────┘
+                                      │
+                                      ▼
+┌─────────────────────────────────────────────────────────────────────────────┐
+│  3. PROMPT INJECTION                                                        │
+│  ───────────────────────────────────────────────────────────────────────── │
+│  # Memories are injected into the system message:                           │
+│                                                                             │
+│  messages = [                                                               │
+│      {"role": "system", "content": """                                      │
+│          # Relevant Memories                                                │
+│          1. [WORLD] User prefers pytest for testing                         │
+│          2. [WORLD] User is building a FastAPI app                          │
+│          3. [OPINION] User likes type hints                                 │
+│      """},                                                                  │
+│      {"role": "user", "content": "Help me with my Python project"}          │
+│  ]                                                                          │
+└─────────────────────────────────────────────────────────────────────────────┘
+                                      │
+                                      ▼
+┌─────────────────────────────────────────────────────────────────────────────┐
+│  4. LLM CALL                                                                │
+│  ───────────────────────────────────────────────────────────────────────── │
+│  # The enriched prompt is sent to the LLM                                   │
+│  response = litellm.completion(model="gpt-4o-mini", messages=messages)      │
+│                                                                             │
+│  # LLM now has context and can give personalized responses like:            │
+│  # "Since you're working on your FastAPI app, here's how to add tests       │
+│  #  with pytest..."                                                         │
+└─────────────────────────────────────────────────────────────────────────────┘
+                                      │
+                                      ▼
+┌─────────────────────────────────────────────────────────────────────────────┐
+│  5. CONVERSATION STORAGE (after LLM call)                                   │
+│  ───────────────────────────────────────────────────────────────────────── │
+│  # The conversation is stored to Hindsight for future recall                │
+│  hindsight.retain(                                                          │
+│      content="User: Help me with my Python project\n"                       │
+│              "Assistant: Since you're working on FastAPI..."                │
+│  )                                                                          │
+│  # Hindsight extracts facts: "User asked about Python project help"         │
+└─────────────────────────────────────────────────────────────────────────────┘
+                                      │
+                                      ▼
+┌─────────────────────────────────────────────────────────────────────────────┐
+│  6. RESPONSE RETURNED                                                       │
+│  ───────────────────────────────────────────────────────────────────────── │
+│  # You receive the response as normal                                       │
+│  print(response.choices[0].message.content)                                 │
+└─────────────────────────────────────────────────────────────────────────────┘
+```
+
+The memory injection and storage happen automatically - you just use `completion()` as normal.
+
+## Configuration Options
+
+```python
+hindsight_litellm.configure(
+    # Required
+    hindsight_api_url="http://localhost:8888",  # Hindsight API server URL
+    bank_id="my-agent",                          # Memory bank ID
+
+    api_key="your-api-key",        # Optional API key for authentication
+
+    # Optional - Memory behavior
+    store_conversations=True,      # Store conversations after LLM calls
+    inject_memories=True,          # Inject relevant memories into prompts
+    use_reflect=False,             # Use reflect API (synthesized) vs recall (raw memories)
+    reflect_include_facts=False,   # Include source facts with reflect responses
+    max_memories=None,             # Maximum memories to inject (None = unlimited)
+    max_memory_tokens=4096,        # Maximum tokens for memory context
+    recall_budget="mid",           # Recall budget: "low", "mid", "high"
+    fact_types=["world", "agent"], # Filter fact types to inject
+
+    # Optional - Bank Configuration
+    bank_name="My Agent",          # Human-readable display name for the memory bank
+    background="This agent...",    # Instructions guiding what Hindsight should remember (see below)
+
+    # Optional - Advanced
+    injection_mode="system_message",  # or "prepend_user"
+    excluded_models=["gpt-3.5*"],     # Exclude certain models
+    verbose=True,                     # Enable verbose logging and debug info
+)
+```
+
+### Bank Configuration: background and bank_name
+
+The `background` and `bank_name` parameters configure the memory bank itself. When provided, `configure()` will automatically create or update the bank with these settings.
+
+- **bank_name**: A human-readable display name for the memory bank. Useful for identifying banks in the Hindsight UI or when managing multiple banks.
+
+- **background**: Instructions that guide Hindsight on what information is important to extract and remember from conversations. This influences memory extraction during the `retain` operation and can affect how the bank's "disposition" (skepticism, literalism, empathy) is calibrated.
+
+```python
+# Example: Customer support routing agent
+hindsight_litellm.configure(
+    hindsight_api_url="http://localhost:8888",
+    bank_id="support-router",
+    bank_name="Customer Support Router",
+    background="""This agent routes customer support requests to the appropriate team.
+    Remember which types of issues should go to which teams (billing, technical, sales).
+    Track customer preferences for communication channels and past issue resolutions.
+    Note any escalation patterns or VIP customers who need special handling.""",
+)
+```
+
+### Memory Modes: Reflect vs Recall
+
+- **Recall mode** (`use_reflect=False`, default): Retrieves raw memory facts and injects them as a numbered list. Best when you need precise, individual memories.
+- **Reflect mode** (`use_reflect=True`): Synthesizes memories into a coherent context paragraph. Best for natural, conversational memory context.
+
+```python
+# Recall mode - raw memories
+hindsight_litellm.configure(
+    bank_id="my-agent",
+    use_reflect=False,  # Default
+)
+# Injects: "1. [WORLD] User prefers Python\n2. [OPINION] User dislikes Java..."
+
+# Reflect mode - synthesized context
+hindsight_litellm.configure(
+    bank_id="my-agent",
+    use_reflect=True,
+)
+# Injects: "Based on previous conversations, the user is a Python developer who..."
+```
+
+## Multi-Provider Support
+
+Works with any LiteLLM-supported provider:
+
+```python
+import hindsight_litellm
+
+hindsight_litellm.configure(
+    hindsight_api_url="http://localhost:8888",
+    bank_id="my-agent",
+)
+hindsight_litellm.enable()
+
+# OpenAI
+hindsight_litellm.completion(model="gpt-4o", messages=[...])
+
+# Anthropic
+hindsight_litellm.completion(model="claude-3-5-sonnet-20241022", messages=[...])
+
+# Groq
+hindsight_litellm.completion(model="groq/llama-3.1-70b-versatile", messages=[...])
+
+# Azure OpenAI
+hindsight_litellm.completion(model="azure/gpt-4", messages=[...])
+
+# AWS Bedrock
+hindsight_litellm.completion(model="bedrock/anthropic.claude-3", messages=[...])
+
+# Google Vertex AI
+hindsight_litellm.completion(model="vertex_ai/gemini-pro", messages=[...])
+```
+
+## Direct Memory APIs
+
+### Recall - Query raw memories
+
+```python
+from hindsight_litellm import configure, recall
+
+configure(bank_id="my-agent", hindsight_api_url="http://localhost:8888")
+
+# Query memories
+memories = recall("what projects am I working on?", budget="mid")
+for m in memories:
+    print(f"- [{m.fact_type}] {m.text}")
+
+# Output:
+# - [world] User is building a FastAPI project
+# - [opinion] User prefers Python over JavaScript
+```
+
+### Reflect - Get synthesized context
+
+```python
+from hindsight_litellm import configure, reflect
+
+configure(bank_id="my-agent", hindsight_api_url="http://localhost:8888")
+
+# Get synthesized memory context
+result = reflect("what do you know about the user's preferences?")
+print(result.text)
+
+# Output:
+# "Based on our conversations, the user prefers Python for backend development..."
+```
+
+### Retain - Store memories
+
+```python
+from hindsight_litellm import configure, retain
+
+configure(bank_id="my-agent", hindsight_api_url="http://localhost:8888")
+
+# Store a memory
+result = retain(
+    content="User mentioned they're working on a machine learning project",
+    context="Discussion about current projects",
+)
+print(f"Retained successfully: {result.success}, items: {result.items_count}")
+```
+
+### Async APIs
+
+```python
+from hindsight_litellm import arecall, areflect, aretain
+
+# Async versions of all memory APIs
+memories = await arecall("what do you know about me?")
+context = await areflect("summarize user preferences")
+result = await aretain(content="New information to remember")
+```
+
+## Native Client Wrappers
+
+Alternative to LiteLLM callbacks for direct SDK integration:
+
+### OpenAI Wrapper
+
+```python
+from openai import OpenAI
+from hindsight_litellm import wrap_openai
+
+client = OpenAI()
+wrapped = wrap_openai(
+    client,
+    bank_id="my-agent",
+    hindsight_api_url="http://localhost:8888",
+)
+
+response = wrapped.chat.completions.create(
+    model="gpt-4",
+    messages=[{"role": "user", "content": "What do you know about me?"}]
+)
+```
+
+### Anthropic Wrapper
+
+```python
+from anthropic import Anthropic
+from hindsight_litellm import wrap_anthropic
+
+client = Anthropic()
+wrapped = wrap_anthropic(
+    client,
+    bank_id="my-agent",
+    hindsight_api_url="http://localhost:8888",
+)
+
+response = wrapped.messages.create(
+    model="claude-3-5-sonnet-20241022",
+    max_tokens=1024,
+    messages=[{"role": "user", "content": "Hello!"}]
+)
+```
+
+## Debug Mode
+
+When `verbose=True`, you can inspect exactly what memories are being injected:
+
+```python
+from hindsight_litellm import configure, enable, completion, get_last_injection_debug
+
+configure(
+    bank_id="my-agent",
+    hindsight_api_url="http://localhost:8888",
+    verbose=True,
+    use_reflect=True,
+)
+enable()
+
+response = completion(
+    model="gpt-4o-mini",
+    messages=[{"role": "user", "content": "What's my favorite color?"}]
+)
+
+# Inspect what was injected
+debug = get_last_injection_debug()
+if debug:
+    print(f"Mode: {debug.mode}")           # "reflect" or "recall"
+    print(f"Injected: {debug.injected}")   # True/False
+    print(f"Results: {debug.results_count}")
+    print(f"Memory context:\n{debug.memory_context}")
+    if debug.error:
+        print(f"Error: {debug.error}")
+```
+
+## Context Manager
+
+```python
+from hindsight_litellm import hindsight_memory
+import litellm
+
+with hindsight_memory(bank_id="user-123"):
+    response = litellm.completion(model="gpt-4", messages=[...])
+# Memory integration automatically disabled after context
+```
+
+## Disabling and Cleanup
+
+```python
+from hindsight_litellm import disable, cleanup
+
+# Temporarily disable memory integration
+disable()
+
+# Clean up all resources (call when shutting down)
+cleanup()
+```
+
+## API Reference
+
+### Main Functions
+
+| Function | Description |
+|----------|-------------|
+| `configure(...)` | Configure global Hindsight settings |
+| `enable()` | Enable memory integration with LiteLLM |
+| `disable()` | Disable memory integration |
+| `is_enabled()` | Check if memory integration is enabled |
+| `cleanup()` | Clean up all resources |
+
+### Configuration Functions
+
+| Function | Description |
+|----------|-------------|
+| `get_config()` | Get current configuration |
+| `is_configured()` | Check if Hindsight is configured |
+| `reset_config()` | Reset configuration to defaults |
+
+### Memory Functions
+
+| Function | Description |
+|----------|-------------|
+| `recall(query, ...)` | Synchronously query raw memories |
+| `arecall(query, ...)` | Asynchronously query raw memories |
+| `reflect(query, ...)` | Synchronously get synthesized memory context |
+| `areflect(query, ...)` | Asynchronously get synthesized memory context |
+| `retain(content, ...)` | Synchronously store a memory |
+| `aretain(content, ...)` | Asynchronously store a memory |
+
+### Debug Functions
+
+| Function | Description |
+|----------|-------------|
+| `get_last_injection_debug()` | Get debug info from last memory injection |
+| `clear_injection_debug()` | Clear stored debug info |
+
+### Client Wrappers
+
+| Function | Description |
+|----------|-------------|
+| `wrap_openai(client, ...)` | Wrap OpenAI client with memory |
+| `wrap_anthropic(client, ...)` | Wrap Anthropic client with memory |
+
+## Requirements
+
+- Python >= 3.10
+- litellm >= 1.40.0
+- A running Hindsight API server
+
+## License
+
+MIT
diff --git a/hindsight-integrations/litellm/hindsight_litellm/__init__.py b/hindsight-integrations/litellm/hindsight_litellm/__init__.py
new file mode 100644
index 00000000..437a54b6
--- /dev/null
+++ b/hindsight-integrations/litellm/hindsight_litellm/__init__.py
@@ -0,0 +1,817 @@
+"""Hindsight-LiteLLM: Universal LLM memory integration via LiteLLM.
+
+This package provides automatic memory integration for any LLM provider
+supported by LiteLLM (100+ providers including OpenAI, Anthropic, Groq,
+Azure, AWS Bedrock, Google Vertex AI, and more).
+
+Features:
+- Automatic memory injection before LLM calls
+- Automatic conversation storage after LLM calls
+- Works with any LiteLLM-supported provider
+- Zero code changes to existing LiteLLM usage
+- Multi-user support via separate bank_ids
+- Document grouping for conversation threading
+- Direct recall API for manual memory queries
+- Native client wrappers for OpenAI and Anthropic
+
+Basic usage:
+    >>> from hindsight_litellm import configure, enable
+    >>>
+    >>> # Configure Hindsight integration
+    >>> configure(
+    ...     hindsight_api_url="http://localhost:8888",
+    ...     bank_id="user-123",  # Use separate bank_ids for multi-user support
+    ...     store_conversations=True,
+    ...     inject_memories=True,
+    ... )
+    >>>
+    >>> # Enable memory integration
+    >>> enable()
+    >>>
+    >>> # Now use LiteLLM as normal - memory integration is automatic
+    >>> import litellm
+    >>> response = litellm.completion(
+    ...     model="gpt-4",
+    ...     messages=[{"role": "user", "content": "What did we discuss about AI?"}]
+    ... )
+
+Direct recall API:
+    >>> from hindsight_litellm import configure, recall
+    >>> configure(bank_id="my-agent", hindsight_api_url="http://localhost:8888")
+    >>>
+    >>> # Query memories directly
+    >>> memories = recall("what projects am I working on?")
+    >>> for m in memories:
+    ...     print(f"- [{m.fact_type}] {m.text}")
+
+Native client wrappers:
+    >>> from openai import OpenAI
+    >>> from hindsight_litellm import wrap_openai
+    >>>
+    >>> client = OpenAI()
+    >>> wrapped = wrap_openai(client, bank_id="user-123")
+    >>>
+    >>> response = wrapped.chat.completions.create(
+    ...     model="gpt-4",
+    ...     messages=[{"role": "user", "content": "Hello!"}]
+    ... )
+
+Works with any LiteLLM-supported provider:
+    >>> # OpenAI
+    >>> litellm.completion(model="gpt-4", messages=[...])
+    >>>
+    >>> # Anthropic
+    >>> litellm.completion(model="claude-3-opus-20240229", messages=[...])
+    >>>
+    >>> # Groq
+    >>> litellm.completion(model="groq/llama-3.1-70b-versatile", messages=[...])
+    >>>
+    >>> # Azure OpenAI
+    >>> litellm.completion(model="azure/gpt-4", messages=[...])
+    >>>
+    >>> # AWS Bedrock
+    >>> litellm.completion(model="bedrock/anthropic.claude-3", messages=[...])
+    >>>
+    >>> # Google Vertex AI
+    >>> litellm.completion(model="vertex_ai/gemini-pro", messages=[...])
+
+Context manager usage:
+    >>> from hindsight_litellm import hindsight_memory
+    >>>
+    >>> with hindsight_memory(bank_id="user-123"):
+    ...     response = litellm.completion(model="gpt-4", messages=[...])
+    >>> # Memory integration automatically disabled after context
+
+Configuration options:
+    - hindsight_api_url: URL of your Hindsight API server
+    - bank_id: Memory bank ID for memory operations (required). For multi-user
+        support, use different bank_ids per user (e.g., f"user-{user_id}")
+    - api_key: Optional API key for Hindsight authentication
+    - store_conversations: Whether to store conversations (default: True)
+    - inject_memories: Whether to inject relevant memories (default: True)
+    - injection_mode: How to inject memories (system_message or prepend_user)
+    - max_memories: Maximum number of memories to inject (None = unlimited)
+    - recall_budget: Budget for memory recall (low, mid, high)
+    - excluded_models: List of model patterns to exclude from interception
+    - verbose: Enable verbose logging
+    - bank_name: Display name for the memory bank
+    - background: Instructions that help Hindsight understand what to remember
+
+Background example:
+    >>> configure(
+    ...     bank_id="routing-agent",
+    ...     background="This agent routes customer requests to support channels. "
+    ...                "Remember which types of issues should go to which channels.",
+    ... )
+"""
+
+from contextlib import contextmanager
+from dataclasses import dataclass
+from typing import Optional, List, Any
+
+import litellm
+
+from .config import (
+    configure,
+    get_config,
+    is_configured,
+    reset_config,
+    HindsightConfig,
+    MemoryInjectionMode,
+)
+from .callbacks import (
+    HindsightCallback,
+    get_callback,
+    cleanup_callback,
+)
+from .wrappers import (
+    recall,
+    arecall,
+    RecallResult,
+    RecallResponse,
+    RecallDebugInfo,
+    reflect,
+    areflect,
+    ReflectResult,
+    ReflectDebugInfo,
+    retain,
+    aretain,
+    RetainResult,
+    RetainDebugInfo,
+    wrap_openai,
+    wrap_anthropic,
+    HindsightOpenAI,
+    HindsightAnthropic,
+)
+
+
+__version__ = "0.1.0"
+
+# Track whether we've registered with LiteLLM
+_enabled = False
+
+# Store original functions for restoration
+_original_completion = None
+_original_acompletion = None
+
+
+@dataclass
+class InjectionDebugInfo:
+    """Debug information from a memory injection operation.
+
+    This is populated when verbose=True in the config and can be retrieved
+    via get_last_injection_debug() after a completion() call.
+
+    Attributes:
+        mode: The injection mode used ("reflect" or "recall")
+        query: The user query used for memory lookup
+        bank_id: The bank ID used
+        memory_context: The formatted memory context that was injected
+        reflect_text: The raw reflect text (when mode="reflect")
+        reflect_facts: The facts used to generate the reflect response (when reflect_include_facts=True)
+        recall_results: The raw recall results (when mode="recall")
+        results_count: Number of memories/results found
+        injected: Whether memories were actually injected into the prompt
+        error: Error message if injection failed (None on success)
+    """
+    mode: str  # "reflect" or "recall"
+    query: str
+    bank_id: str
+    memory_context: str  # The formatted context that was injected
+    reflect_text: Optional[str] = None  # Raw reflect response text
+    reflect_facts: Optional[List[dict]] = None  # Facts used by reflect (when reflect_include_facts=True)
+    recall_results: Optional[List[dict]] = None  # Raw recall results
+    results_count: int = 0
+    injected: bool = False
+    error: Optional[str] = None  # Error message if injection failed
+
+
+# Store the last injection debug info (populated when verbose=True)
+_last_injection_debug: Optional[InjectionDebugInfo] = None
+
+
+def get_last_injection_debug() -> Optional[InjectionDebugInfo]:
+    """Get debug info from the last memory injection operation.
+
+    When verbose=True in the config, this returns information about
+    what memories were injected into the last completion() call.
+
+    Returns:
+        InjectionDebugInfo if verbose mode captured injection info, None otherwise
+
+    Example:
+        >>> from hindsight_litellm import configure, enable, completion, get_last_injection_debug
+        >>> configure(bank_id="my-agent", verbose=True, use_reflect=True)
+        >>> enable()
+        >>> response = completion(model="gpt-4o-mini", messages=[...])
+        >>> debug = get_last_injection_debug()
+        >>> if debug:
+        ...     print(f"Injected {debug.results_count} memories via {debug.mode}")
+        ...     print(f"Reflect text: {debug.reflect_text}")
+    """
+    return _last_injection_debug
+
+
+def clear_injection_debug() -> None:
+    """Clear the stored injection debug info."""
+    global _last_injection_debug
+    _last_injection_debug = None
+
+
+def _inject_memories(messages: List[dict]) -> List[dict]:
+    """Inject memories into messages list.
+
+    Returns the modified messages list with memories injected into the system message.
+    Uses reflect API when config.use_reflect=True, otherwise uses recall API.
+
+    When verbose=True in config, stores debug info retrievable via get_last_injection_debug().
+    """
+    global _last_injection_debug
+    import logging
+
+    # Clear previous debug info
+    _last_injection_debug = None
+
+    if not is_configured():
+        return messages
+
+    config = get_config()
+    if not config or not config.enabled or not config.inject_memories:
+        return messages
+
+    if not messages:
+        return messages
+
+    # Extract user query from last user message
+    user_query = None
+    for msg in reversed(messages):
+        if msg.get("role") == "user":
+            content = msg.get("content")
+            if isinstance(content, str):
+                user_query = content
+                break
+
+    if not user_query:
+        return messages
+
+    try:
+        from hindsight_client import Hindsight
+
+        # Use bank_id directly (no entity scoping)
+        bank_id = config.bank_id
+
+        # Track debug info
+        mode = "reflect" if config.use_reflect else "recall"
+        reflect_text = None
+        reflect_facts = None
+        recall_results = None
+        results_count = 0
+        memory_context = ""
+
+        # Create client
+        client = Hindsight(base_url=config.hindsight_api_url, timeout=30.0)
+
+        # Use reflect API if use_reflect is enabled
+        if config.use_reflect:
+            # If reflect_include_facts is enabled, use the API directly to include facts
+            if config.reflect_include_facts:
+                from hindsight_client_api.models import reflect_request, reflect_include_options
+                request_obj = reflect_request.ReflectRequest(
+                    query=user_query,
+                    budget=config.recall_budget or "mid",
+                    include=reflect_include_options.ReflectIncludeOptions(facts={}),
+                )
+                import asyncio
+                try:
+                    loop = asyncio.get_event_loop()
+                except RuntimeError:
+                    loop = asyncio.new_event_loop()
+                    asyncio.set_event_loop(loop)
+                result = loop.run_until_complete(client._api.reflect(bank_id, request_obj))
+                # Extract facts from based_on
+                if hasattr(result, 'based_on') and result.based_on:
+                    reflect_facts = [
+                        {
+                            "text": f.text if hasattr(f, 'text') else str(f),
+                            "type": getattr(f, 'type', None),
+                            "context": getattr(f, 'context', None),
+                        }
+                        for f in result.based_on
+                    ]
+            else:
+                result = client.reflect(
+                    bank_id=bank_id,
+                    query=user_query,
+                    budget=config.recall_budget or "mid",
+                )
+            reflect_text = result.text if hasattr(result, 'text') else str(result)
+
+            if not reflect_text:
+                # Store debug info for empty result
+                if config.verbose:
+                    _last_injection_debug = InjectionDebugInfo(
+                        mode=mode,
+                        query=user_query,
+                        bank_id=bank_id,
+                        memory_context="",
+                        reflect_text="",
+                        reflect_facts=reflect_facts,
+                        results_count=0,
+                        injected=False,
+                    )
+                return messages
+
+            results_count = 1  # reflect returns a single synthesized response
+            memory_context = (
+                "# Relevant Context from Memory\n"
+                f"{reflect_text}"
+            )
+        else:
+            # Use recall API (original behavior)
+            result = client.recall(
+                bank_id=bank_id,
+                query=user_query,
+                budget=config.recall_budget or "mid",
+                max_tokens=config.max_memory_tokens or 4096,
+                types=config.fact_types,
+            )
+            # client.recall() returns a list directly, not an object with .results
+            if isinstance(result, list):
+                results = result
+            elif hasattr(result, 'results'):
+                results = result.results
+            else:
+                results = []
+            # Convert to dicts for debug info
+            recall_results = [
+                {
+                    "text": r.text if hasattr(r, 'text') else str(r),
+                    "type": getattr(r, 'type', 'world'),
+                }
+                for r in results
+            ]
+
+            if not results:
+                # Store debug info for empty result
+                if config.verbose:
+                    _last_injection_debug = InjectionDebugInfo(
+                        mode=mode,
+                        query=user_query,
+                        bank_id=bank_id,
+                        memory_context="",
+                        recall_results=[],
+                        results_count=0,
+                        injected=False,
+                    )
+                return messages
+
+            # Format memories (apply limit if set, otherwise use all)
+            results_to_use = results[:config.max_memories] if config.max_memories else results
+            memory_lines = []
+            for i, r in enumerate(results_to_use, 1):
+                text = r.text if hasattr(r, 'text') else str(r)
+                fact_type = getattr(r, 'type', 'world')
+                if text:
+                    type_label = fact_type.upper() if fact_type else "MEMORY"
+                    memory_lines.append(f"{i}. [{type_label}] {text}")
+
+            if not memory_lines:
+                if config.verbose:
+                    _last_injection_debug = InjectionDebugInfo(
+                        mode=mode,
+                        query=user_query,
+                        bank_id=bank_id,
+                        memory_context="",
+                        recall_results=recall_results,
+                        results_count=0,
+                        injected=False,
+                    )
+                return messages
+
+            results_count = len(memory_lines)
+            memory_context = (
+                "# Relevant Memories\n"
+                "The following information from memory may be relevant:\n\n"
+                + "\n".join(memory_lines)
+            )
+
+        # Inject into messages
+        updated_messages = list(messages)
+
+        # Find existing system message or create new one
+        found_system = False
+        for i, msg in enumerate(updated_messages):
+            if msg.get("role") == "system":
+                existing_content = msg.get("content", "")
+                updated_messages[i] = {
+                    **msg,
+                    "content": f"{existing_content}\n\n{memory_context}"
+                }
+                found_system = True
+                break
+
+        if not found_system:
+            updated_messages.insert(0, {
+                "role": "system",
+                "content": memory_context
+            })
+
+        # Store debug info when verbose
+        if config.verbose:
+            _last_injection_debug = InjectionDebugInfo(
+                mode=mode,
+                query=user_query,
+                bank_id=bank_id,
+                memory_context=memory_context,
+                reflect_text=reflect_text,
+                reflect_facts=reflect_facts,
+                recall_results=recall_results,
+                results_count=results_count,
+                injected=True,
+            )
+            logger = logging.getLogger("hindsight_litellm")
+            logger.info(f"Injected memories using {mode} into prompt")
+
+        return updated_messages
+
+    except ImportError as e:
+        if config.verbose:
+            logging.getLogger("hindsight_litellm").warning(
+                f"hindsight_client not installed: {e}. Install with: pip install hindsight-client"
+            )
+            _last_injection_debug = InjectionDebugInfo(
+                mode="reflect" if config.use_reflect else "recall",
+                query=user_query or "",
+                bank_id=config.bank_id or "",
+                memory_context="",
+                results_count=0,
+                injected=False,
+                error=f"hindsight_client not installed: {e}",
+            )
+        return messages
+    except Exception as e:
+        # Always set debug info on error when verbose mode is on
+        if config.verbose:
+            logging.getLogger("hindsight_litellm").warning(f"Failed to inject memories: {e}")
+            _last_injection_debug = InjectionDebugInfo(
+                mode="reflect" if config.use_reflect else "recall",
+                query=user_query or "",
+                bank_id=config.bank_id or "",
+                memory_context="",
+                results_count=0,
+                injected=False,
+                error=str(e),
+            )
+        return messages
+
+
+def _wrapped_completion(*args, **kwargs):
+    """Wrapper for litellm.completion that injects memories before the call."""
+    # Inject memories into messages
+    if "messages" in kwargs:
+        kwargs["messages"] = _inject_memories(kwargs["messages"])
+    elif args and len(args) > 1:
+        # messages might be second positional arg after model
+        args = list(args)
+        if isinstance(args[1], list):
+            args[1] = _inject_memories(args[1])
+        args = tuple(args)
+
+    # Call original
+    return _original_completion(*args, **kwargs)
+
+
+async def _wrapped_acompletion(*args, **kwargs):
+    """Wrapper for litellm.acompletion that injects memories before the call."""
+    # Inject memories into messages
+    if "messages" in kwargs:
+        kwargs["messages"] = _inject_memories(kwargs["messages"])
+    elif args and len(args) > 1:
+        args = list(args)
+        if isinstance(args[1], list):
+            args[1] = _inject_memories(args[1])
+        args = tuple(args)
+
+    # Call original
+    return await _original_acompletion(*args, **kwargs)
+
+
+def enable() -> None:
+    """Enable Hindsight memory integration with LiteLLM.
+
+    This monkeypatches LiteLLM functions to:
+    1. Inject relevant memories into prompts before LLM calls
+    2. Store conversations to Hindsight after successful LLM calls
+
+    Must be called after configure() to take effect.
+
+    Example:
+        >>> from hindsight_litellm import configure, enable
+        >>> configure(bank_id="my-agent", hindsight_api_url="http://localhost:8888")
+        >>> enable()
+        >>>
+        >>> # Now all LiteLLM calls will have memory integration
+        >>> import litellm
+        >>> response = litellm.completion(model="gpt-4", messages=[...])
+    """
+    global _enabled, _original_completion, _original_acompletion
+
+    if _enabled:
+        return  # Already enabled
+
+    if not is_configured():
+        raise RuntimeError(
+            "Hindsight not configured. Call configure() before enable()."
+        )
+
+    # Store original functions and monkeypatch for memory injection
+    _original_completion = litellm.completion
+    _original_acompletion = litellm.acompletion
+    litellm.completion = _wrapped_completion
+    litellm.acompletion = _wrapped_acompletion
+
+    # Get or create the callback instance for storing conversations
+    callback = get_callback()
+
+    # Register callback using litellm.callbacks for conversation storage
+    if callback not in litellm.callbacks:
+        litellm.callbacks.append(callback)
+
+    _enabled = True
+
+    config = get_config()
+    if config and config.verbose:
+        print(f"Hindsight memory enabled for bank: {config.bank_id}")
+
+
+def disable() -> None:
+    """Disable Hindsight memory integration with LiteLLM.
+
+    This restores the original LiteLLM functions and removes callbacks,
+    stopping memory injection and conversation storage.
+
+    Example:
+        >>> from hindsight_litellm import disable
+        >>> disable()  # Stop memory integration
+    """
+    global _enabled, _original_completion, _original_acompletion
+
+    if not _enabled:
+        return  # Already disabled
+
+    # Restore original functions
+    if _original_completion is not None:
+        litellm.completion = _original_completion
+        _original_completion = None
+    if _original_acompletion is not None:
+        litellm.acompletion = _original_acompletion
+        _original_acompletion = None
+
+    # Remove callback from litellm.callbacks
+    callback = get_callback()
+    if callback in litellm.callbacks:
+        litellm.callbacks.remove(callback)
+
+    _enabled = False
+
+    config = get_config()
+    if config and config.verbose:
+        print("Hindsight memory disabled")
+
+
+def is_enabled() -> bool:
+    """Check if Hindsight memory integration is currently enabled.
+
+    Returns:
+        True if enable() has been called and not subsequently disabled
+    """
+    return _enabled
+
+
+def cleanup() -> None:
+    """Clean up all Hindsight resources.
+
+    This disables the integration and closes any open connections.
+    Call this when shutting down your application.
+
+    Example:
+        >>> from hindsight_litellm import cleanup
+        >>> cleanup()  # Clean up when done
+    """
+    disable()
+    cleanup_callback()
+    reset_config()
+
+
+# =============================================================================
+# Convenience wrappers - use hindsight_litellm.completion() directly
+# =============================================================================
+
+def completion(*args, **kwargs):
+    """Call LiteLLM completion with Hindsight memory integration.
+
+    This is a convenience wrapper that delegates to litellm.completion().
+    Memory injection and storage happen automatically if configured and enabled.
+
+    Args:
+        *args: Positional arguments passed to litellm.completion()
+        **kwargs: Keyword arguments passed to litellm.completion()
+
+    Returns:
+        LiteLLM ModelResponse object
+
+    Example:
+        >>> import hindsight_litellm
+        >>>
+        >>> hindsight_litellm.configure(
+        ...     hindsight_api_url="http://localhost:8888",
+        ...     bank_id="my-agent",
+        ... )
+        >>> hindsight_litellm.enable()
+        >>>
+        >>> # Use directly - no need to import litellm separately
+        >>> response = hindsight_litellm.completion(
+        ...     model="gpt-4o-mini",
+        ...     messages=[{"role": "user", "content": "Hello!"}]
+        ... )
+    """
+    return litellm.completion(*args, **kwargs)
+
+
+async def acompletion(*args, **kwargs):
+    """Call LiteLLM async completion with Hindsight memory integration.
+
+    This is a convenience wrapper that delegates to litellm.acompletion().
+    Memory injection and storage happen automatically if configured and enabled.
+
+    Args:
+        *args: Positional arguments passed to litellm.acompletion()
+        **kwargs: Keyword arguments passed to litellm.acompletion()
+
+    Returns:
+        LiteLLM ModelResponse object
+
+    Example:
+        >>> import hindsight_litellm
+        >>> import asyncio
+        >>>
+        >>> hindsight_litellm.configure(
+        ...     hindsight_api_url="http://localhost:8888",
+        ...     bank_id="my-agent",
+        ... )
+        >>> hindsight_litellm.enable()
+        >>>
+        >>> async def main():
+        ...     response = await hindsight_litellm.acompletion(
+        ...         model="gpt-4o-mini",
+        ...         messages=[{"role": "user", "content": "Hello!"}]
+        ...     )
+        ...     return response
+        >>>
+        >>> asyncio.run(main())
+    """
+    return await litellm.acompletion(*args, **kwargs)
+
+
+@contextmanager
+def hindsight_memory(
+    hindsight_api_url: str = "http://localhost:8888",
+    bank_id: Optional[str] = None,
+    api_key: Optional[str] = None,
+    store_conversations: bool = True,
+    inject_memories: bool = True,
+    injection_mode: MemoryInjectionMode = MemoryInjectionMode.SYSTEM_MESSAGE,
+    max_memories: Optional[int] = None,
+    max_memory_tokens: int = 4096,
+    recall_budget: str = "mid",
+    fact_types: Optional[List[str]] = None,
+    document_id: Optional[str] = None,
+    excluded_models: Optional[List[str]] = None,
+    verbose: bool = False,
+    bank_name: Optional[str] = None,
+    background: Optional[str] = None,
+):
+    """Context manager for temporary Hindsight memory integration.
+
+    Use this to enable memory integration for a specific block of code,
+    automatically cleaning up afterwards.
+
+    Args:
+        hindsight_api_url: URL of the Hindsight API server
+        bank_id: Memory bank ID for memory operations (required). For multi-user
+            support, use different bank_ids per user (e.g., f"user-{user_id}")
+        api_key: Optional API key for Hindsight authentication
+        store_conversations: Whether to store conversations
+        inject_memories: Whether to inject relevant memories
+        injection_mode: How to inject memories
+        max_memories: Maximum number of memories to inject (None = unlimited)
+        max_memory_tokens: Maximum tokens for memory context
+        recall_budget: Budget for memory recall (low, mid, high)
+        fact_types: List of fact types to filter (world, agent, opinion, observation)
+        document_id: Optional document ID for grouping conversations
+        excluded_models: List of model patterns to exclude
+        verbose: Enable verbose logging
+        bank_name: Optional display name for the memory bank
+        background: Optional background/instructions for memory extraction
+
+    Example:
+        >>> from hindsight_litellm import hindsight_memory
+        >>> import litellm
+        >>>
+        >>> with hindsight_memory(bank_id="user-123"):
+        ...     response = litellm.completion(model="gpt-4", messages=[...])
+        >>> # Memory integration automatically disabled after context
+    """
+    # Save previous state
+    was_enabled = is_enabled()
+    previous_config = get_config()
+
+    try:
+        # Configure and enable
+        configure(
+            hindsight_api_url=hindsight_api_url,
+            bank_id=bank_id,
+            api_key=api_key,
+            store_conversations=store_conversations,
+            inject_memories=inject_memories,
+            injection_mode=injection_mode,
+            max_memories=max_memories,
+            max_memory_tokens=max_memory_tokens,
+            recall_budget=recall_budget,
+            fact_types=fact_types,
+            document_id=document_id,
+            excluded_models=excluded_models,
+            verbose=verbose,
+            bank_name=bank_name,
+            background=background,
+        )
+        enable()
+        yield
+    finally:
+        # Restore previous state
+        disable()
+        if previous_config:
+            configure(
+                hindsight_api_url=previous_config.hindsight_api_url,
+                bank_id=previous_config.bank_id,
+                api_key=previous_config.api_key,
+                store_conversations=previous_config.store_conversations,
+                inject_memories=previous_config.inject_memories,
+                injection_mode=previous_config.injection_mode,
+                max_memories=previous_config.max_memories,
+                max_memory_tokens=previous_config.max_memory_tokens,
+                recall_budget=previous_config.recall_budget,
+                fact_types=previous_config.fact_types,
+                document_id=previous_config.document_id,
+                excluded_models=previous_config.excluded_models,
+                verbose=previous_config.verbose,
+                bank_name=previous_config.bank_name,
+                background=previous_config.background,
+            )
+            if was_enabled:
+                enable()
+        else:
+            reset_config()
+
+
+__all__ = [
+    # Main API
+    "configure",
+    "enable",
+    "disable",
+    "is_enabled",
+    "cleanup",
+    "hindsight_memory",
+    # LLM completion wrappers (convenience)
+    "completion",
+    "acompletion",
+    # Direct memory APIs
+    "recall",
+    "arecall",
+    "RecallResult",
+    "reflect",
+    "areflect",
+    "ReflectResult",
+    "retain",
+    "aretain",
+    "RetainResult",
+    # Native client wrappers
+    "wrap_openai",
+    "wrap_anthropic",
+    "HindsightOpenAI",
+    "HindsightAnthropic",
+    # Configuration
+    "get_config",
+    "is_configured",
+    "reset_config",
+    "HindsightConfig",
+    "MemoryInjectionMode",
+    # Injection debug (verbose mode)
+    "get_last_injection_debug",
+    "clear_injection_debug",
+    "InjectionDebugInfo",
+    # Callback (for advanced usage)
+    "HindsightCallback",
+    "get_callback",
+    "cleanup_callback",
+]
diff --git a/hindsight-integrations/litellm/hindsight_litellm/callbacks.py b/hindsight-integrations/litellm/hindsight_litellm/callbacks.py
new file mode 100644
index 00000000..f6478593
--- /dev/null
+++ b/hindsight-integrations/litellm/hindsight_litellm/callbacks.py
@@ -0,0 +1,640 @@
+"""LiteLLM callback handlers for Hindsight memory integration.
+
+This module implements LiteLLM's CustomLogger interface to intercept
+LLM calls and integrate with Hindsight for memory injection and storage.
+
+Uses direct HTTP calls via requests/httpx to avoid async event loop conflicts
+when the hindsight_client's async methods are called from LiteLLM callbacks.
+"""
+
+import logging
+import fnmatch
+import hashlib
+from datetime import datetime, timezone
+from typing import Any, Dict, List, Optional, Set
+import asyncio
+import threading
+import concurrent.futures
+
+from litellm.integrations.custom_logger import CustomLogger
+from litellm.types.utils import ModelResponse
+
+from .config import get_config, is_configured, HindsightConfig, MemoryInjectionMode
+
+# Use requests for sync HTTP calls to avoid async event loop issues
+try:
+    import requests
+    HAS_REQUESTS = True
+except ImportError:
+    HAS_REQUESTS = False
+
+try:
+    import httpx
+    HAS_HTTPX = True
+except ImportError:
+    HAS_HTTPX = False
+
+
+logger = logging.getLogger(__name__)
+
+# Thread pool for running async operations in background
+_executor = concurrent.futures.ThreadPoolExecutor(max_workers=4, thread_name_prefix="hindsight-")
+
+
+class HindsightCallback(CustomLogger):
+    """LiteLLM custom logger that integrates with Hindsight memory system.
+
+    This callback handler:
+    1. Injects relevant memories into prompts before LLM calls
+    2. Stores conversations to Hindsight after successful LLM calls
+
+    Features:
+    - Works with 100+ LLM providers via LiteLLM
+    - Deduplication to avoid storing duplicate conversations
+    - Configurable memory injection modes
+    - Support for entity observations in recall
+
+    Usage:
+        >>> from hindsight_litellm import configure, enable
+        >>> configure(bank_id="my-agent", hindsight_api_url="http://localhost:8888")
+        >>> enable()
+        >>>
+        >>> # Now all LiteLLM calls will have memory integration
+        >>> import litellm
+        >>> response = litellm.completion(
+        ...     model="gpt-4",
+        ...     messages=[{"role": "user", "content": "What did we discuss?"}]
+        ... )
+    """
+
+    def __init__(self):
+        """Initialize the Hindsight callback handler."""
+        super().__init__()
+        self._http_session = None
+        self._http_lock = threading.Lock()
+        # Track recently stored conversation hashes for deduplication
+        self._recent_hashes: Set[str] = set()
+        self._max_hash_cache = 1000
+
+    def _get_http_session(self):
+        """Get or create a requests Session (thread-safe)."""
+        if self._http_session is None:
+            with self._http_lock:
+                if self._http_session is None:
+                    if HAS_REQUESTS:
+                        self._http_session = requests.Session()
+                    elif HAS_HTTPX:
+                        self._http_session = httpx.Client(timeout=30.0)
+                    else:
+                        raise RuntimeError(
+                            "Neither 'requests' nor 'httpx' is installed. "
+                            "Please install one: pip install requests"
+                        )
+        return self._http_session
+
+    def _http_post(self, url: str, json_data: dict, config: HindsightConfig) -> Optional[dict]:
+        """Make a synchronous HTTP POST request."""
+        try:
+            session = self._get_http_session()
+            headers = {"Content-Type": "application/json"}
+            if config.api_key:
+                headers["Authorization"] = f"Bearer {config.api_key}"
+
+            if HAS_REQUESTS:
+                response = session.post(url, json=json_data, headers=headers, timeout=30)
+                response.raise_for_status()
+                return response.json()
+            elif HAS_HTTPX:
+                response = session.post(url, json=json_data, headers=headers)
+                response.raise_for_status()
+                return response.json()
+        except Exception as e:
+            if config.verbose:
+                logger.warning(f"HTTP POST failed: {e}")
+            return None
+
+    def _should_skip_model(self, model: str, config: HindsightConfig) -> bool:
+        """Check if this model should be excluded from interception."""
+        for pattern in config.excluded_models:
+            if fnmatch.fnmatch(model.lower(), pattern.lower()):
+                return True
+        return False
+
+    def _extract_user_query(self, messages: List[Dict[str, Any]]) -> Optional[str]:
+        """Extract the user's query from the last user message."""
+        for msg in reversed(messages):
+            role = msg.get("role", "")
+            if role == "user":
+                content = msg.get("content")
+                if isinstance(content, str):
+                    return content
+                elif isinstance(content, list):
+                    # Handle structured content (e.g., vision messages)
+                    text_parts = []
+                    for item in content:
+                        if isinstance(item, dict) and item.get("type") == "text":
+                            text_parts.append(item.get("text", ""))
+                    if text_parts:
+                        return " ".join(text_parts)
+        return None
+
+    def _compute_conversation_hash(
+        self,
+        user_input: str,
+        assistant_output: str,
+    ) -> str:
+        """Compute a hash for deduplication."""
+        content = f"{user_input.strip().lower()}|{assistant_output.strip().lower()}"
+        return hashlib.md5(content.encode()).hexdigest()[:16]
+
+    def _is_duplicate(self, conv_hash: str) -> bool:
+        """Check if this conversation was recently stored."""
+        if conv_hash in self._recent_hashes:
+            return True
+
+        # Add to cache, evict oldest if full
+        self._recent_hashes.add(conv_hash)
+        if len(self._recent_hashes) > self._max_hash_cache:
+            # Remove oldest (arbitrary since set, but good enough)
+            self._recent_hashes.pop()
+
+        return False
+
+    def _format_memories(
+        self,
+        results: List[Any],
+        config: HindsightConfig
+    ) -> str:
+        """Format memory recall results into a context string.
+
+        Results can be RecallResult objects (with .text, .type attributes)
+        or dicts (with get() method).
+        """
+        if not results:
+            return ""
+
+        # Apply limit if set, otherwise use all results
+        results_to_use = results[:config.max_memories] if config.max_memories else results
+        memory_lines = []
+        for i, result in enumerate(results_to_use, 1):
+            # Handle both RecallResult objects and dicts
+            if hasattr(result, 'text'):
+                text = result.text or ""
+                fact_type = getattr(result, 'type', 'world') or "world"
+                weight = getattr(result, 'weight', 0.0) or 0.0
+            else:
+                text = result.get("text", "")
+                fact_type = result.get("type", result.get("fact_type", "world"))
+                weight = result.get("weight", 0.0)
+
+            if text:
+                # Include metadata for context
+                type_label = fact_type.upper() if fact_type else "MEMORY"
+                line = f"{i}. [{type_label}] {text}"
+                if weight > 0 and config.verbose:
+                    line += f" (relevance: {weight:.2f})"
+                memory_lines.append(line)
+
+        if not memory_lines:
+            return ""
+
+        return (
+            "# Relevant Memories\n"
+            "The following information from memory may be relevant:\n\n"
+            + "\n".join(memory_lines)
+        )
+
+    def _inject_memories_into_messages(
+        self,
+        messages: List[Dict[str, Any]],
+        memory_context: str,
+        config: HindsightConfig,
+    ) -> List[Dict[str, Any]]:
+        """Inject memory context into the messages list."""
+        if not memory_context:
+            return messages
+
+        updated_messages = list(messages)  # Make a copy
+
+        if config.injection_mode == MemoryInjectionMode.SYSTEM_MESSAGE:
+            # Find existing system message or create new one
+            for i, msg in enumerate(updated_messages):
+                if msg.get("role") == "system":
+                    # Append to existing system message
+                    existing_content = msg.get("content", "")
+                    updated_messages[i] = {
+                        **msg,
+                        "content": f"{existing_content}\n\n{memory_context}"
+                    }
+                    return updated_messages
+
+            # No system message found, prepend one
+            updated_messages.insert(0, {
+                "role": "system",
+                "content": memory_context
+            })
+
+        elif config.injection_mode == MemoryInjectionMode.PREPEND_USER:
+            # Find the last user message and prepend context
+            for i in range(len(updated_messages) - 1, -1, -1):
+                if updated_messages[i].get("role") == "user":
+                    original_content = updated_messages[i].get("content", "")
+                    if isinstance(original_content, str):
+                        updated_messages[i] = {
+                            **updated_messages[i],
+                            "content": f"{memory_context}\n\n---\n\n{original_content}"
+                        }
+                    break
+
+        return updated_messages
+
+    def _get_bank_id(self, config: HindsightConfig) -> str:
+        """Get the bank_id for API calls."""
+        return config.bank_id
+
+    def _recall_memories_sync(
+        self,
+        query: str,
+        config: HindsightConfig
+    ) -> List[Dict[str, Any]]:
+        """Recall relevant memories from Hindsight (sync) using direct HTTP."""
+        try:
+            bank_id = self._get_bank_id(config)
+            url = f"{config.hindsight_api_url}/v1/default/banks/{bank_id}/memories/recall"
+
+            request_data = {
+                "query": query,
+                "budget": config.recall_budget or "mid",
+                "max_tokens": config.max_memory_tokens or 4096,
+            }
+            if config.fact_types:
+                request_data["types"] = config.fact_types
+
+            response = self._http_post(url, request_data, config)
+            if response and "results" in response:
+                return response["results"]
+            return []
+
+        except Exception as e:
+            if config.verbose:
+                logger.warning(f"Failed to recall memories: {e}")
+            return []
+
+    async def _recall_memories_async(
+        self,
+        query: str,
+        config: HindsightConfig
+    ) -> List[Any]:
+        """Recall relevant memories from Hindsight (async).
+
+        Uses thread pool executor with sync HTTP to avoid event loop conflicts.
+        """
+        try:
+            loop = asyncio.get_running_loop()
+            results = await loop.run_in_executor(
+                _executor,
+                self._recall_memories_sync,
+                query,
+                config
+            )
+
+            return results if isinstance(results, list) else []
+
+        except Exception as e:
+            if config.verbose:
+                logger.warning(f"Failed to recall memories: {e}")
+            return []
+
+    def _store_conversation_sync(
+        self,
+        messages: List[Dict[str, Any]],
+        response: ModelResponse,
+        model: str,
+        config: HindsightConfig,
+    ) -> None:
+        """Store the conversation to Hindsight (sync) using direct HTTP.
+
+        By default, stores the full conversation history passed to the LLM.
+        Each message is stored as a separate item, all linked by document_id.
+
+        Hindsight will process the document as a whole for memory extraction.
+        """
+        try:
+            # Extract assistant response from the LLM response
+            assistant_output = ""
+            if response.choices and len(response.choices) > 0:
+                choice = response.choices[0]
+                if hasattr(choice, "message") and choice.message:
+                    assistant_output = choice.message.content or ""
+
+            if not assistant_output:
+                return
+
+            # Build conversation items - each message becomes a separate item
+            # All linked by document_id for Hindsight to process together
+            items = []
+            for msg in messages:
+                role = msg.get("role", "").upper()
+                content = msg.get("content", "")
+
+                # Skip system messages - they're instructions, not conversation
+                if role == "SYSTEM":
+                    continue
+
+                # Skip if this looks like our injected memory context
+                if isinstance(content, str) and content.startswith("# Relevant Memories"):
+                    continue
+
+                # Handle structured content (e.g., vision messages)
+                if isinstance(content, list):
+                    text_parts = []
+                    for item in content:
+                        if isinstance(item, dict) and item.get("type") == "text":
+                            text_parts.append(item.get("text", ""))
+                    content = " ".join(text_parts)
+
+                if content:
+                    # Map roles to clearer labels
+                    label = "USER" if role == "USER" else "ASSISTANT"
+                    items.append(f"{label}: {content}")
+
+            # Add the new assistant response
+            items.append(f"ASSISTANT: {assistant_output}")
+
+            if not items:
+                return
+
+            # Use last user message for deduplication hash
+            user_input = self._extract_user_query(messages) or ""
+
+            # Deduplication check
+            conv_hash = self._compute_conversation_hash(user_input, assistant_output)
+            if self._is_duplicate(conv_hash):
+                if config.verbose:
+                    logger.debug(f"Skipping duplicate conversation: {conv_hash}")
+                return
+
+            # Build the full conversation as a single item for now
+            # (Future: could store each message as separate item in same document)
+            conversation_text = "\n\n".join(items)
+
+            # Build metadata
+            metadata = {
+                "source": "litellm",
+                "model": model,
+            }
+
+            # Add token usage if available
+            if hasattr(response, "usage") and response.usage:
+                if hasattr(response.usage, "total_tokens"):
+                    metadata["tokens"] = str(response.usage.total_tokens)
+
+            bank_id = self._get_bank_id(config)
+            url = f"{config.hindsight_api_url}/v1/default/banks/{bank_id}/memories"
+
+            request_data = {
+                "items": [
+                    {
+                        "content": conversation_text,
+                        "context": f"conversation:litellm:{model}",
+                        "metadata": metadata,
+                        "document_id": config.document_id,  # Group by document
+                    }
+                ],
+            }
+
+            self._http_post(url, request_data, config)
+
+            if config.verbose:
+                logger.info(f"Stored conversation to Hindsight bank: {config.bank_id}")
+
+        except Exception as e:
+            if config.verbose:
+                logger.warning(f"Failed to store conversation: {e}")
+
+    async def _store_conversation_async(
+        self,
+        messages: List[Dict[str, Any]],
+        response: ModelResponse,
+        model: str,
+        config: HindsightConfig,
+    ) -> None:
+        """Store the conversation to Hindsight (async).
+
+        Uses thread pool executor with sync HTTP to avoid event loop conflicts.
+        """
+        try:
+            loop = asyncio.get_running_loop()
+            await loop.run_in_executor(
+                _executor,
+                self._store_conversation_sync,
+                messages,
+                response,
+                model,
+                config
+            )
+        except Exception as e:
+            if config.verbose:
+                logger.warning(f"Failed to store conversation: {e}")
+
+    # ========== LiteLLM CustomLogger Interface ==========
+
+    def log_pre_api_call(
+        self,
+        model: str,
+        messages: List[Dict[str, Any]],
+        kwargs: Dict[str, Any],
+    ) -> None:
+        """Called before making the API call (sync).
+
+        This is where we inject memories into the messages.
+        """
+        if not is_configured():
+            return
+
+        config = get_config()
+        if not config or not config.enabled or not config.inject_memories:
+            return
+
+        if self._should_skip_model(model, config):
+            return
+
+        # Extract user query
+        user_query = self._extract_user_query(messages)
+        if not user_query:
+            return
+
+        # Recall relevant memories
+        memories = self._recall_memories_sync(user_query, config)
+        if not memories:
+            return
+
+        # Format and inject memories
+        memory_context = self._format_memories(memories, config)
+        updated_messages = self._inject_memories_into_messages(
+            messages, memory_context, config
+        )
+
+        # Modify messages list IN-PLACE (don't just reassign kwargs)
+        messages.clear()
+        messages.extend(updated_messages)
+
+        if config.verbose:
+            logger.info(f"Injected {len(memories)} memories into prompt")
+
+    async def async_log_pre_api_call(
+        self,
+        model: str,
+        messages: List[Dict[str, Any]],
+        kwargs: Dict[str, Any],
+    ) -> None:
+        """Called before making the API call (async).
+
+        This is where we inject memories into the messages.
+        """
+        if not is_configured():
+            return
+
+        config = get_config()
+        if not config or not config.enabled or not config.inject_memories:
+            return
+
+        if self._should_skip_model(model, config):
+            return
+
+        # Extract user query
+        user_query = self._extract_user_query(messages)
+        if not user_query:
+            return
+
+        # Recall relevant memories
+        memories = await self._recall_memories_async(user_query, config)
+        if not memories:
+            return
+
+        # Format and inject memories
+        memory_context = self._format_memories(memories, config)
+        updated_messages = self._inject_memories_into_messages(
+            messages, memory_context, config
+        )
+
+        # Modify messages list IN-PLACE (don't just reassign kwargs)
+        messages.clear()
+        messages.extend(updated_messages)
+
+        if config.verbose:
+            logger.info(f"Injected {len(memories)} memories into prompt")
+
+    def log_success_event(
+        self,
+        kwargs: Dict[str, Any],
+        response_obj: Any,
+        start_time: float,
+        end_time: float,
+    ) -> None:
+        """Called after successful API call (sync).
+
+        This is where we store the conversation.
+        """
+        if not is_configured():
+            return
+
+        config = get_config()
+        if not config or not config.enabled or not config.store_conversations:
+            return
+
+        model = kwargs.get("model", "unknown")
+        if self._should_skip_model(model, config):
+            return
+
+        messages = kwargs.get("messages", [])
+        if not messages:
+            return
+
+        # Store the conversation
+        self._store_conversation_sync(messages, response_obj, model, config)
+
+    async def async_log_success_event(
+        self,
+        kwargs: Dict[str, Any],
+        response_obj: Any,
+        start_time: float,
+        end_time: float,
+    ) -> None:
+        """Called after successful API call (async).
+
+        This is where we store the conversation.
+        """
+        if not is_configured():
+            return
+
+        config = get_config()
+        if not config or not config.enabled or not config.store_conversations:
+            return
+
+        model = kwargs.get("model", "unknown")
+        if self._should_skip_model(model, config):
+            return
+
+        messages = kwargs.get("messages", [])
+        if not messages:
+            return
+
+        # Store the conversation
+        await self._store_conversation_async(messages, response_obj, model, config)
+
+    def log_failure_event(
+        self,
+        kwargs: Dict[str, Any],
+        response_obj: Any,
+        start_time: float,
+        end_time: float,
+    ) -> None:
+        """Called after failed API call (sync)."""
+        # We don't store failed conversations
+        pass
+
+    async def async_log_failure_event(
+        self,
+        kwargs: Dict[str, Any],
+        response_obj: Any,
+        start_time: float,
+        end_time: float,
+    ) -> None:
+        """Called after failed API call (async)."""
+        # We don't store failed conversations
+        pass
+
+    def close(self) -> None:
+        """Clean up resources."""
+        with self._http_lock:
+            if self._http_session is not None:
+                try:
+                    if HAS_REQUESTS:
+                        self._http_session.close()
+                    elif HAS_HTTPX:
+                        self._http_session.close()
+                except Exception:
+                    pass
+                self._http_session = None
+        self._recent_hashes.clear()
+
+
+# Global callback instance
+_callback: Optional[HindsightCallback] = None
+
+
+def get_callback() -> HindsightCallback:
+    """Get the global callback instance, creating it if necessary."""
+    global _callback
+    if _callback is None:
+        _callback = HindsightCallback()
+    return _callback
+
+
+def cleanup_callback() -> None:
+    """Clean up the global callback instance."""
+    global _callback
+    if _callback is not None:
+        _callback.close()
+        _callback = None
diff --git a/hindsight-integrations/litellm/hindsight_litellm/config.py b/hindsight-integrations/litellm/hindsight_litellm/config.py
new file mode 100644
index 00000000..06673977
--- /dev/null
+++ b/hindsight-integrations/litellm/hindsight_litellm/config.py
@@ -0,0 +1,232 @@
+"""Global configuration for Hindsight-LiteLLM integration."""
+
+from typing import Optional, List
+from dataclasses import dataclass, field
+from enum import Enum
+
+
+class MemoryInjectionMode(str, Enum):
+    """How memories should be injected into the prompt."""
+    SYSTEM_MESSAGE = "system_message"  # Add as system message
+    PREPEND_USER = "prepend_user"  # Prepend to user message
+    DISABLED = "disabled"  # Don't inject memories
+
+
+@dataclass
+class HindsightConfig:
+    """Configuration for Hindsight integration with LiteLLM.
+
+    Attributes:
+        hindsight_api_url: URL of the Hindsight API server
+        bank_id: Memory bank ID for memory operations (required). For multi-user
+            support, use different bank_ids per user (e.g., f"user-{user_id}")
+        api_key: Optional API key for Hindsight authentication
+        store_conversations: Whether to store conversations to Hindsight
+        inject_memories: Whether to inject relevant memories into prompts
+        injection_mode: How to inject memories (system_message or prepend_user)
+        max_memories: Maximum number of memories to inject
+        max_memory_tokens: Maximum tokens for injected memory context
+        recall_budget: Budget level for memory recall (low, mid, high)
+        fact_types: List of fact types to filter recall (world, agent, opinion, observation)
+        document_id: Optional document ID for grouping stored conversations
+        enabled: Master switch to enable/disable Hindsight integration
+        excluded_models: List of model patterns to exclude from interception
+        verbose: Enable verbose logging
+        bank_name: Optional display name for the memory bank
+        background: Optional background/instructions for memory extraction
+        use_reflect: Use reflect API instead of recall for memory injection (synthesizes answer)
+    """
+
+    hindsight_api_url: str = "http://localhost:8888"
+    bank_id: Optional[str] = None
+    api_key: Optional[str] = None
+    store_conversations: bool = True
+    inject_memories: bool = True
+    injection_mode: MemoryInjectionMode = MemoryInjectionMode.SYSTEM_MESSAGE
+    max_memories: Optional[int] = None  # None = no limit (use all results from API)
+    max_memory_tokens: int = 4096
+    recall_budget: str = "mid"  # low, mid, high
+    fact_types: Optional[List[str]] = None  # world, agent, opinion, observation
+    document_id: Optional[str] = None
+    enabled: bool = True
+    excluded_models: List[str] = field(default_factory=list)
+    verbose: bool = False
+    bank_name: Optional[str] = None  # Display name for the memory bank
+    background: Optional[str] = None  # Background/instructions for memory extraction
+    use_reflect: bool = False  # Use reflect instead of recall for memory injection
+    reflect_include_facts: bool = False  # Include facts used by reflect in debug info
+
+
+# Global configuration instance
+_global_config: Optional[HindsightConfig] = None
+
+
+def configure(
+    hindsight_api_url: str = "http://localhost:8888",
+    bank_id: Optional[str] = None,
+    api_key: Optional[str] = None,
+    store_conversations: bool = True,
+    inject_memories: bool = True,
+    injection_mode: MemoryInjectionMode = MemoryInjectionMode.SYSTEM_MESSAGE,
+    max_memories: Optional[int] = None,
+    max_memory_tokens: int = 4096,
+    recall_budget: str = "mid",
+    fact_types: Optional[List[str]] = None,
+    document_id: Optional[str] = None,
+    enabled: bool = True,
+    excluded_models: Optional[List[str]] = None,
+    verbose: bool = False,
+    bank_name: Optional[str] = None,
+    background: Optional[str] = None,
+    use_reflect: bool = False,
+    reflect_include_facts: bool = False,
+) -> HindsightConfig:
+    """Configure global Hindsight integration settings for LiteLLM.
+
+    This function sets up the global configuration that will be used by the
+    LiteLLM callbacks to inject memories and store conversations.
+
+    Args:
+        hindsight_api_url: URL of the Hindsight API server
+        bank_id: Memory bank ID for memory operations (required). For multi-user
+            support, use different bank_ids per user (e.g., f"user-{user_id}")
+        api_key: Optional API key for Hindsight authentication
+        store_conversations: Whether to store conversations to Hindsight
+        inject_memories: Whether to inject relevant memories into prompts
+        injection_mode: How to inject memories into the prompt
+        max_memories: Maximum number of memories to inject
+        max_memory_tokens: Maximum tokens for injected memory context
+        recall_budget: Budget level for memory recall (low, mid, high)
+        fact_types: List of fact types to filter (world, agent, opinion, observation)
+        document_id: Optional document ID for grouping stored conversations
+        enabled: Master switch to enable/disable Hindsight integration
+        excluded_models: List of model patterns to exclude from interception
+        verbose: Enable verbose logging
+        bank_name: Optional display name for the memory bank
+        background: Optional background/instructions that help Hindsight understand
+            what information is important to extract and remember from conversations.
+            This is passed to create_bank() to configure the memory bank.
+        use_reflect: Use reflect API instead of recall for memory injection.
+            When True, Hindsight will synthesize a contextual answer based on
+            memories rather than returning raw memory facts.
+        reflect_include_facts: When use_reflect=True, include the facts that
+            were used to generate the reflect response in the debug info.
+            This is useful for debugging what memories the reflect API used.
+
+    Returns:
+        The configured HindsightConfig instance
+
+    Example:
+        >>> from hindsight_litellm import configure, enable
+        >>> configure(
+        ...     hindsight_api_url="http://localhost:8888",
+        ...     bank_id="user-123",  # Per-user bank for multi-user support
+        ...     store_conversations=True,
+        ...     inject_memories=True,
+        ...     background="This agent routes customer requests to support channels. "
+        ...                "Remember which types of issues should go to which channels.",
+        ... )
+        >>> enable()  # Register callbacks with LiteLLM
+    """
+    global _global_config
+
+    _global_config = HindsightConfig(
+        hindsight_api_url=hindsight_api_url,
+        bank_id=bank_id,
+        api_key=api_key,
+        store_conversations=store_conversations,
+        inject_memories=inject_memories,
+        injection_mode=injection_mode,
+        max_memories=max_memories,
+        max_memory_tokens=max_memory_tokens,
+        recall_budget=recall_budget,
+        fact_types=fact_types,
+        document_id=document_id,
+        enabled=enabled,
+        excluded_models=excluded_models or [],
+        verbose=verbose,
+        bank_name=bank_name,
+        background=background,
+        use_reflect=use_reflect,
+        reflect_include_facts=reflect_include_facts,
+    )
+
+    # If background or bank_name is provided, create/update the bank
+    if bank_id and (background or bank_name):
+        _create_or_update_bank(
+            hindsight_api_url=hindsight_api_url,
+            bank_id=bank_id,
+            name=bank_name,
+            background=background,
+            verbose=verbose,
+        )
+
+    return _global_config
+
+
+def _create_or_update_bank(
+    hindsight_api_url: str,
+    bank_id: str,
+    name: Optional[str] = None,
+    background: Optional[str] = None,
+    verbose: bool = False,
+) -> None:
+    """Create or update a memory bank with the given configuration.
+
+    This is called automatically by configure() when background or bank_name is provided.
+    """
+    try:
+        from hindsight_client import Hindsight
+
+        client = Hindsight(hindsight_api_url)
+        client.create_bank(
+            bank_id=bank_id,
+            name=name,
+            background=background,
+        )
+        if verbose:
+            import logging
+            logging.getLogger("hindsight_litellm").info(
+                f"Created/updated bank '{bank_id}' with background"
+            )
+    except ImportError:
+        if verbose:
+            import logging
+            logging.getLogger("hindsight_litellm").warning(
+                "hindsight_client not installed. Cannot create bank with background. "
+                "Install with: pip install hindsight-client"
+            )
+    except Exception as e:
+        if verbose:
+            import logging
+            logging.getLogger("hindsight_litellm").warning(
+                f"Failed to create/update bank: {e}"
+            )
+
+
+def get_config() -> Optional[HindsightConfig]:
+    """Get the current global configuration.
+
+    Returns:
+        The current HindsightConfig instance, or None if not configured
+    """
+    return _global_config
+
+
+def is_configured() -> bool:
+    """Check if Hindsight has been configured.
+
+    Returns:
+        True if configure() has been called with a valid bank_id
+    """
+    return (
+        _global_config is not None
+        and _global_config.enabled
+        and _global_config.bank_id is not None
+    )
+
+
+def reset_config() -> None:
+    """Reset the global configuration to None."""
+    global _global_config
+    _global_config = None
diff --git a/hindsight-integrations/litellm/hindsight_litellm/wrappers.py b/hindsight-integrations/litellm/hindsight_litellm/wrappers.py
new file mode 100644
index 00000000..9b5feecc
--- /dev/null
+++ b/hindsight-integrations/litellm/hindsight_litellm/wrappers.py
@@ -0,0 +1,1000 @@
+"""Native client wrappers for Hindsight memory integration.
+
+This module provides wrappers for native LLM client SDKs (OpenAI, Anthropic)
+that automatically integrate with Hindsight for memory injection and storage.
+
+This is an alternative to the LiteLLM callback approach, providing direct
+integration with native client libraries.
+"""
+
+import logging
+from typing import Any, Dict, List, Optional, Union
+from dataclasses import dataclass
+
+from .config import get_config, is_configured, HindsightConfig
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class RecallResult:
+    """A single memory recall result."""
+    text: str
+    fact_type: str
+    weight: float
+    metadata: Optional[Dict[str, Any]] = None
+
+    def __str__(self) -> str:
+        return self.text
+
+
+@dataclass
+class RecallDebugInfo:
+    """Debug information from a recall operation."""
+    query: str
+    bank_id: str
+    budget: str
+    max_tokens: int
+    fact_types: Optional[List[str]]
+    results_count: int
+    api_url: str
+
+
+@dataclass
+class RecallResponse:
+    """Response from a recall operation, including results and optional debug info."""
+    results: List[RecallResult]
+    debug: Optional[RecallDebugInfo] = None
+
+    def __iter__(self):
+        return iter(self.results)
+
+    def __len__(self):
+        return len(self.results)
+
+    def __getitem__(self, key):
+        return self.results[key]
+
+    def __bool__(self):
+        return bool(self.results)
+
+
+def recall(
+    query: str,
+    bank_id: Optional[str] = None,
+    fact_types: Optional[List[str]] = None,
+    budget: Optional[str] = None,
+    max_tokens: Optional[int] = None,
+    hindsight_api_url: Optional[str] = None,
+) -> RecallResponse:
+    """Recall memories from Hindsight.
+
+    This function allows you to manually query memories without making an LLM call.
+    Useful for debugging, building custom UIs, or pre-filtering memories.
+
+    Args:
+        query: The query string to search memories for
+        bank_id: Override the configured bank_id. For multi-user support,
+            use different bank_ids per user (e.g., f"user-{user_id}")
+        fact_types: Filter by fact types (world, agent, opinion, observation)
+        budget: Recall budget level (low, mid, high) - controls how many memories are returned
+        max_tokens: Maximum tokens for memory context
+        hindsight_api_url: Override the configured API URL
+
+    Returns:
+        RecallResponse containing matched memories (iterable like a list).
+        When verbose=True in config, includes debug info via .debug attribute.
+
+    Raises:
+        RuntimeError: If Hindsight is not configured and no overrides provided
+
+    Example:
+        >>> from hindsight_litellm import configure, recall
+        >>> configure(bank_id="my-agent", hindsight_api_url="http://localhost:8888")
+        >>>
+        >>> # Query memories
+        >>> memories = recall("what projects am I working on?")
+        >>> for m in memories:
+        ...     print(f"- [{m.fact_type}] {m.text}")
+        - [world] User is building a FastAPI project
+        - [opinion] User prefers Python over JavaScript
+        >>>
+        >>> # With verbose mode, access debug info
+        >>> configure(bank_id="my-agent", verbose=True)
+        >>> memories = recall("what projects am I working on?")
+        >>> if memories.debug:
+        ...     print(f"Queried bank: {memories.debug.bank_id}")
+    """
+    # Get config or use overrides
+    config = get_config()
+
+    api_url = hindsight_api_url or (config.hindsight_api_url if config else None)
+    target_bank_id = bank_id or (config.bank_id if config else None)
+    target_fact_types = fact_types or (config.fact_types if config else None)
+    target_budget = budget or (config.recall_budget if config else "mid")
+    target_max_tokens = max_tokens or (config.max_memory_tokens if config else 4096)
+
+    if not api_url or not target_bank_id:
+        raise RuntimeError(
+            "Hindsight not configured. Call configure() or provide bank_id and hindsight_api_url."
+        )
+
+    try:
+        from hindsight_client import Hindsight
+
+        client = Hindsight(base_url=api_url, timeout=30.0)
+
+        # Call recall API
+        results = client.recall(
+            bank_id=target_bank_id,
+            query=query,
+            types=target_fact_types,
+            budget=target_budget,
+            max_tokens=target_max_tokens,
+        )
+
+        # Convert to RecallResult objects
+        recall_results = []
+        if results:
+            for r in results:
+                if hasattr(r, 'text'):
+                    # Object with attributes
+                    fact_type = getattr(r, 'type', None) or getattr(r, 'fact_type', 'unknown')
+                    recall_results.append(RecallResult(
+                        text=r.text,
+                        fact_type=fact_type,
+                        weight=getattr(r, 'weight', 0.0),
+                        metadata=getattr(r, 'metadata', None),
+                    ))
+                elif isinstance(r, dict):
+                    # Dict from API response - API returns 'type' not 'fact_type'
+                    fact_type = r.get('type') or r.get('fact_type', 'unknown')
+                    recall_results.append(RecallResult(
+                        text=r.get('text', str(r)),
+                        fact_type=fact_type,
+                        weight=r.get('weight', 0.0),
+                        metadata=r.get('metadata'),
+                    ))
+
+        # Include debug info if verbose
+        debug_info = None
+        if config and config.verbose:
+            debug_info = RecallDebugInfo(
+                query=query,
+                bank_id=target_bank_id,
+                budget=target_budget,
+                max_tokens=target_max_tokens,
+                fact_types=target_fact_types,
+                results_count=len(recall_results),
+                api_url=api_url,
+            )
+
+        return RecallResponse(results=recall_results, debug=debug_info)
+
+    except ImportError as e:
+        raise RuntimeError(f"hindsight-client not installed: {e}")
+    except Exception as e:
+        if config and config.verbose:
+            logger.warning(f"Failed to recall memories: {e}")
+        raise
+
+
+async def arecall(
+    query: str,
+    bank_id: Optional[str] = None,
+    fact_types: Optional[List[str]] = None,
+    budget: Optional[str] = None,
+    max_tokens: Optional[int] = None,
+    hindsight_api_url: Optional[str] = None,
+) -> RecallResponse:
+    """Async version of recall().
+
+    See recall() for full documentation.
+    """
+    import asyncio
+    loop = asyncio.get_event_loop()
+    return await loop.run_in_executor(
+        None,
+        lambda: recall(
+            query=query,
+            bank_id=bank_id,
+            fact_types=fact_types,
+            budget=budget,
+            max_tokens=max_tokens,
+            hindsight_api_url=hindsight_api_url,
+        )
+    )
+
+
+@dataclass
+class ReflectDebugInfo:
+    """Debug information from a reflect operation."""
+    query: str
+    bank_id: str
+    budget: str
+    context: Optional[str]
+    api_url: str
+
+
+@dataclass
+class ReflectResult:
+    """Result from a reflect operation."""
+    text: str
+    based_on: Optional[Dict[str, List[Any]]] = None
+    debug: Optional[ReflectDebugInfo] = None
+
+    def __str__(self) -> str:
+        return self.text
+
+
+def reflect(
+    query: str,
+    bank_id: Optional[str] = None,
+    budget: Optional[str] = None,
+    context: Optional[str] = None,
+    hindsight_api_url: Optional[str] = None,
+) -> ReflectResult:
+    """Generate a contextual answer based on memories.
+
+    Unlike recall() which returns raw memory facts, reflect() uses an LLM
+    to synthesize a coherent answer based on the bank's memories.
+
+    Args:
+        query: The question or prompt to answer
+        bank_id: Override the configured bank_id. For multi-user support,
+            use different bank_ids per user (e.g., f"user-{user_id}")
+        budget: Budget level for reflection (low, mid, high)
+        context: Additional context to include in the reflection
+        hindsight_api_url: Override the configured API URL
+
+    Returns:
+        ReflectResult with synthesized answer text
+
+    Raises:
+        RuntimeError: If Hindsight is not configured and no overrides provided
+
+    Example:
+        >>> from hindsight_litellm import configure, reflect
+        >>> configure(bank_id="my-agent", hindsight_api_url="http://localhost:8888")
+        >>>
+        >>> # Get a synthesized answer based on memories
+        >>> result = reflect("What projects am I working on?")
+        >>> print(result.text)
+        Based on our conversations, you're working on a FastAPI project...
+    """
+    config = get_config()
+
+    api_url = hindsight_api_url or (config.hindsight_api_url if config else None)
+    target_bank_id = bank_id or (config.bank_id if config else None)
+    target_budget = budget or (config.recall_budget if config else "mid")
+
+    if not api_url or not target_bank_id:
+        raise RuntimeError(
+            "Hindsight not configured. Call configure() or provide bank_id and hindsight_api_url."
+        )
+
+    try:
+        from hindsight_client import Hindsight
+
+        client = Hindsight(base_url=api_url, timeout=30.0)
+
+        # Call reflect API
+        result = client.reflect(
+            bank_id=target_bank_id,
+            query=query,
+            budget=target_budget,
+            context=context,
+        )
+
+        # Convert to ReflectResult
+        text = result.text if hasattr(result, 'text') else str(result)
+        based_on = getattr(result, 'based_on', None)
+
+        # Include debug info if verbose
+        debug_info = None
+        if config and config.verbose:
+            debug_info = ReflectDebugInfo(
+                query=query,
+                bank_id=target_bank_id,
+                budget=target_budget,
+                context=context,
+                api_url=api_url,
+            )
+
+        return ReflectResult(text=text, based_on=based_on, debug=debug_info)
+
+    except ImportError as e:
+        raise RuntimeError(f"hindsight-client not installed: {e}")
+    except Exception as e:
+        if config and config.verbose:
+            logger.warning(f"Failed to reflect: {e}")
+        raise
+
+
+async def areflect(
+    query: str,
+    bank_id: Optional[str] = None,
+    budget: Optional[str] = None,
+    context: Optional[str] = None,
+    hindsight_api_url: Optional[str] = None,
+) -> ReflectResult:
+    """Async version of reflect().
+
+    See reflect() for full documentation.
+    """
+    import asyncio
+    loop = asyncio.get_event_loop()
+    return await loop.run_in_executor(
+        None,
+        lambda: reflect(
+            query=query,
+            bank_id=bank_id,
+            budget=budget,
+            context=context,
+            hindsight_api_url=hindsight_api_url,
+        )
+    )
+
+
+@dataclass
+class RetainDebugInfo:
+    """Debug information from a retain operation."""
+    content: str
+    bank_id: str
+    context: Optional[str]
+    document_id: Optional[str]
+    metadata: Optional[Dict[str, str]]
+    api_url: str
+
+
+@dataclass
+class RetainResult:
+    """Result from a retain operation."""
+    success: bool
+    items_count: int = 0
+    debug: Optional[RetainDebugInfo] = None
+
+    def __bool__(self) -> bool:
+        return self.success
+
+
+def retain(
+    content: str,
+    bank_id: Optional[str] = None,
+    context: Optional[str] = None,
+    document_id: Optional[str] = None,
+    metadata: Optional[Dict[str, str]] = None,
+    hindsight_api_url: Optional[str] = None,
+) -> RetainResult:
+    """Store content to Hindsight memory.
+
+    This function allows you to manually store content to memory without
+    making an LLM call. Useful for storing feedback, user preferences,
+    or any other information you want the system to remember.
+
+    Args:
+        content: The text content to store
+        bank_id: Override the configured bank_id. For multi-user support,
+            use different bank_ids per user (e.g., f"user-{user_id}")
+        context: Context description for the memory (e.g., "customer_feedback")
+        document_id: Optional document ID for grouping related memories
+        metadata: Optional key-value metadata to attach to the memory
+        hindsight_api_url: Override the configured API URL
+
+    Returns:
+        RetainResult indicating success
+
+    Raises:
+        RuntimeError: If Hindsight is not configured and no overrides provided
+
+    Example:
+        >>> from hindsight_litellm import configure, retain
+        >>> configure(bank_id="my-agent", hindsight_api_url="http://localhost:8888")
+        >>>
+        >>> # Store feedback
+        >>> retain("User prefers dark mode", context="user_preference")
+        >>>
+        >>> # Store with metadata
+        >>> retain(
+        ...     "Customer reported billing issue resolved",
+        ...     context="support_ticket",
+        ...     metadata={"ticket_id": "12345", "status": "resolved"}
+        ... )
+    """
+    config = get_config()
+
+    api_url = hindsight_api_url or (config.hindsight_api_url if config else None)
+    target_bank_id = bank_id or (config.bank_id if config else None)
+    target_document_id = document_id or (config.document_id if config else None)
+
+    if not api_url or not target_bank_id:
+        raise RuntimeError(
+            "Hindsight not configured. Call configure() or provide bank_id and hindsight_api_url."
+        )
+
+    try:
+        from hindsight_client import Hindsight
+
+        client = Hindsight(base_url=api_url, timeout=30.0)
+
+        # Call retain API
+        result = client.retain(
+            bank_id=target_bank_id,
+            content=content,
+            context=context,
+            document_id=target_document_id,
+            metadata=metadata,
+        )
+
+        # Check success
+        success = getattr(result, 'success', True)
+        items_count = getattr(result, 'items_count', 1)
+
+        # Include debug info if verbose
+        debug_info = None
+        if config and config.verbose:
+            logger.info(f"Stored content to Hindsight bank: {target_bank_id}")
+            debug_info = RetainDebugInfo(
+                content=content,
+                bank_id=target_bank_id,
+                context=context,
+                document_id=target_document_id,
+                metadata=metadata,
+                api_url=api_url,
+            )
+
+        return RetainResult(success=success, items_count=items_count, debug=debug_info)
+
+    except ImportError as e:
+        raise RuntimeError(f"hindsight-client not installed: {e}")
+    except Exception as e:
+        if config and config.verbose:
+            logger.warning(f"Failed to retain: {e}")
+        raise
+
+
+async def aretain(
+    content: str,
+    bank_id: Optional[str] = None,
+    context: Optional[str] = None,
+    document_id: Optional[str] = None,
+    metadata: Optional[Dict[str, str]] = None,
+    hindsight_api_url: Optional[str] = None,
+) -> RetainResult:
+    """Async version of retain().
+
+    See retain() for full documentation.
+    """
+    import asyncio
+    loop = asyncio.get_event_loop()
+    return await loop.run_in_executor(
+        None,
+        lambda: retain(
+            content=content,
+            bank_id=bank_id,
+            context=context,
+            document_id=document_id,
+            metadata=metadata,
+            hindsight_api_url=hindsight_api_url,
+        )
+    )
+
+
+class HindsightOpenAI:
+    """Wrapper for OpenAI client with Hindsight memory integration.
+
+    This wraps the native OpenAI client to automatically inject memories
+    and store conversations.
+
+    Example:
+        >>> from openai import OpenAI
+        >>> from hindsight_litellm import wrap_openai
+        >>>
+        >>> client = OpenAI()
+        >>> wrapped = wrap_openai(client, bank_id="my-agent")
+        >>>
+        >>> response = wrapped.chat.completions.create(
+        ...     model="gpt-4",
+        ...     messages=[{"role": "user", "content": "What do you know about me?"}]
+        ... )
+    """
+
+    def __init__(
+        self,
+        client: Any,
+        bank_id: str,
+        hindsight_api_url: str = "http://localhost:8888",
+        session_id: Optional[str] = None,
+        store_conversations: bool = True,
+        inject_memories: bool = True,
+        max_memories: Optional[int] = None,
+        recall_budget: str = "mid",
+        verbose: bool = False,
+    ):
+        """Initialize the wrapped OpenAI client.
+
+        Args:
+            client: The OpenAI client instance to wrap
+            bank_id: Memory bank ID for memory operations. For multi-user support,
+                use different bank_ids per user (e.g., f"user-{user_id}")
+            hindsight_api_url: URL of the Hindsight API server
+            session_id: Session identifier for conversation grouping
+            store_conversations: Whether to store conversations
+            inject_memories: Whether to inject relevant memories
+            max_memories: Maximum number of memories to inject (None = no limit)
+            recall_budget: Budget level for memory recall (low, mid, high)
+            verbose: Enable verbose logging
+        """
+        self._client = client
+        self._bank_id = bank_id
+        self._api_url = hindsight_api_url
+        self._session_id = session_id
+        self._store_conversations = store_conversations
+        self._inject_memories = inject_memories
+        self._max_memories = max_memories
+        self._recall_budget = recall_budget
+        self._verbose = verbose
+        self._hindsight_client = None
+
+        # Create wrapped chat.completions interface
+        self.chat = _WrappedChat(self)
+
+    def _get_hindsight_client(self):
+        """Get or create the Hindsight client."""
+        if self._hindsight_client is None:
+            from hindsight_client import Hindsight
+            self._hindsight_client = Hindsight(
+                base_url=self._api_url,
+                timeout=30.0,
+            )
+        return self._hindsight_client
+
+    def _recall_memories(self, query: str) -> str:
+        """Recall and format memories for injection."""
+        if not self._inject_memories:
+            return ""
+
+        try:
+            client = self._get_hindsight_client()
+            results = client.recall(
+                bank_id=self._bank_id,
+                query=query,
+                budget=self._recall_budget,
+                max_tokens=self._max_memories * 200 if self._max_memories else 4096,
+            )
+
+            if not results:
+                return ""
+
+            results_to_use = results[:self._max_memories] if self._max_memories else results
+            memory_lines = []
+            for i, r in enumerate(results_to_use, 1):
+                text = r.text if hasattr(r, 'text') else str(r)
+                fact_type = r.fact_type if hasattr(r, 'fact_type') else 'memory'
+                memory_lines.append(f"{i}. [{fact_type.upper()}] {text}")
+
+            if not memory_lines:
+                return ""
+
+            return (
+                "# Relevant Memories\n"
+                "The following information from memory may be relevant:\n\n"
+                + "\n".join(memory_lines)
+            )
+
+        except Exception as e:
+            if self._verbose:
+                logger.warning(f"Failed to recall memories: {e}")
+            return ""
+
+    def _store_conversation(self, user_input: str, assistant_output: str, model: str):
+        """Store the conversation to Hindsight."""
+        if not self._store_conversations:
+            return
+
+        try:
+            client = self._get_hindsight_client()
+            conversation_text = f"USER: {user_input}\n\nASSISTANT: {assistant_output}"
+
+            metadata = {
+                "source": "openai-wrapper",
+                "model": model,
+            }
+            if self._session_id:
+                metadata["session_id"] = self._session_id
+
+            client.retain(
+                bank_id=self._bank_id,
+                content=conversation_text,
+                context=f"conversation:openai:{model}",
+                metadata=metadata,
+            )
+
+            if self._verbose:
+                logger.info(f"Stored conversation to Hindsight")
+
+        except Exception as e:
+            if self._verbose:
+                logger.warning(f"Failed to store conversation: {e}")
+
+    # Proxy other attributes to the underlying client
+    def __getattr__(self, name: str) -> Any:
+        return getattr(self._client, name)
+
+
+class _WrappedChat:
+    """Wrapped chat interface for OpenAI client."""
+
+    def __init__(self, wrapper: HindsightOpenAI):
+        self._wrapper = wrapper
+        self.completions = _WrappedCompletions(wrapper)
+
+
+class _WrappedCompletions:
+    """Wrapped completions interface for OpenAI client."""
+
+    def __init__(self, wrapper: HindsightOpenAI):
+        self._wrapper = wrapper
+
+    def create(self, **kwargs) -> Any:
+        """Create a chat completion with memory integration."""
+        messages = list(kwargs.get("messages", []))
+        model = kwargs.get("model", "gpt-4")
+
+        # Extract user query
+        user_query = None
+        for msg in reversed(messages):
+            if msg.get("role") == "user":
+                content = msg.get("content")
+                if isinstance(content, str):
+                    user_query = content
+                    break
+
+        # Inject memories
+        if user_query and self._wrapper._inject_memories:
+            memory_context = self._wrapper._recall_memories(user_query)
+            if memory_context:
+                # Find system message and append, or prepend new one
+                found_system = False
+                for i, msg in enumerate(messages):
+                    if msg.get("role") == "system":
+                        messages[i] = {
+                            **msg,
+                            "content": f"{msg.get('content', '')}\n\n{memory_context}"
+                        }
+                        found_system = True
+                        break
+
+                if not found_system:
+                    messages.insert(0, {"role": "system", "content": memory_context})
+
+                kwargs["messages"] = messages
+
+        # Make the actual API call
+        response = self._wrapper._client.chat.completions.create(**kwargs)
+
+        # Store conversation
+        if user_query and self._wrapper._store_conversations:
+            if response.choices and response.choices[0].message:
+                assistant_output = response.choices[0].message.content or ""
+                if assistant_output:
+                    self._wrapper._store_conversation(user_query, assistant_output, model)
+
+        return response
+
+
+class HindsightAnthropic:
+    """Wrapper for Anthropic client with Hindsight memory integration.
+
+    This wraps the native Anthropic client to automatically inject memories
+    and store conversations.
+
+    Example:
+        >>> from anthropic import Anthropic
+        >>> from hindsight_litellm import wrap_anthropic
+        >>>
+        >>> client = Anthropic()
+        >>> wrapped = wrap_anthropic(client, bank_id="my-agent")
+        >>>
+        >>> response = wrapped.messages.create(
+        ...     model="claude-3-5-sonnet-20241022",
+        ...     max_tokens=1024,
+        ...     messages=[{"role": "user", "content": "What do you know about me?"}]
+        ... )
+    """
+
+    def __init__(
+        self,
+        client: Any,
+        bank_id: str,
+        hindsight_api_url: str = "http://localhost:8888",
+        session_id: Optional[str] = None,
+        store_conversations: bool = True,
+        inject_memories: bool = True,
+        max_memories: Optional[int] = None,
+        recall_budget: str = "mid",
+        verbose: bool = False,
+    ):
+        """Initialize the wrapped Anthropic client.
+
+        Args:
+            client: The Anthropic client instance to wrap
+            bank_id: Memory bank ID for memory operations. For multi-user support,
+                use different bank_ids per user (e.g., f"user-{user_id}")
+            hindsight_api_url: URL of the Hindsight API server
+            session_id: Session identifier for conversation grouping
+            store_conversations: Whether to store conversations
+            inject_memories: Whether to inject relevant memories
+            max_memories: Maximum number of memories to inject (None = no limit)
+            recall_budget: Budget level for memory recall (low, mid, high)
+            verbose: Enable verbose logging
+        """
+        self._client = client
+        self._bank_id = bank_id
+        self._api_url = hindsight_api_url
+        self._session_id = session_id
+        self._store_conversations = store_conversations
+        self._inject_memories = inject_memories
+        self._max_memories = max_memories
+        self._recall_budget = recall_budget
+        self._verbose = verbose
+        self._hindsight_client = None
+
+        # Create wrapped messages interface
+        self.messages = _WrappedAnthropicMessages(self)
+
+    def _get_hindsight_client(self):
+        """Get or create the Hindsight client."""
+        if self._hindsight_client is None:
+            from hindsight_client import Hindsight
+            self._hindsight_client = Hindsight(
+                base_url=self._api_url,
+                timeout=30.0,
+            )
+        return self._hindsight_client
+
+    def _recall_memories(self, query: str) -> str:
+        """Recall and format memories for injection."""
+        if not self._inject_memories:
+            return ""
+
+        try:
+            client = self._get_hindsight_client()
+            results = client.recall(
+                bank_id=self._bank_id,
+                query=query,
+                budget=self._recall_budget,
+                max_tokens=self._max_memories * 200 if self._max_memories else 4096,
+            )
+
+            if not results:
+                return ""
+
+            results_to_use = results[:self._max_memories] if self._max_memories else results
+            memory_lines = []
+            for i, r in enumerate(results_to_use, 1):
+                text = r.text if hasattr(r, 'text') else str(r)
+                fact_type = r.fact_type if hasattr(r, 'fact_type') else 'memory'
+                memory_lines.append(f"{i}. [{fact_type.upper()}] {text}")
+
+            if not memory_lines:
+                return ""
+
+            return (
+                "# Relevant Memories\n"
+                "The following information from memory may be relevant:\n\n"
+                + "\n".join(memory_lines)
+            )
+
+        except Exception as e:
+            if self._verbose:
+                logger.warning(f"Failed to recall memories: {e}")
+            return ""
+
+    def _store_conversation(self, user_input: str, assistant_output: str, model: str):
+        """Store the conversation to Hindsight."""
+        if not self._store_conversations:
+            return
+
+        try:
+            client = self._get_hindsight_client()
+            conversation_text = f"USER: {user_input}\n\nASSISTANT: {assistant_output}"
+
+            metadata = {
+                "source": "anthropic-wrapper",
+                "model": model,
+            }
+            if self._session_id:
+                metadata["session_id"] = self._session_id
+
+            client.retain(
+                bank_id=self._bank_id,
+                content=conversation_text,
+                context=f"conversation:anthropic:{model}",
+                metadata=metadata,
+            )
+
+            if self._verbose:
+                logger.info(f"Stored conversation to Hindsight")
+
+        except Exception as e:
+            if self._verbose:
+                logger.warning(f"Failed to store conversation: {e}")
+
+    # Proxy other attributes to the underlying client
+    def __getattr__(self, name: str) -> Any:
+        return getattr(self._client, name)
+
+
+class _WrappedAnthropicMessages:
+    """Wrapped messages interface for Anthropic client."""
+
+    def __init__(self, wrapper: HindsightAnthropic):
+        self._wrapper = wrapper
+
+    def create(self, **kwargs) -> Any:
+        """Create a message with memory integration."""
+        messages = list(kwargs.get("messages", []))
+        model = kwargs.get("model", "claude-3-5-sonnet-20241022")
+        system = kwargs.get("system", "")
+
+        # Extract user query
+        user_query = None
+        for msg in reversed(messages):
+            if msg.get("role") == "user":
+                content = msg.get("content")
+                if isinstance(content, str):
+                    user_query = content
+                    break
+                elif isinstance(content, list):
+                    # Handle structured content
+                    for item in content:
+                        if isinstance(item, dict) and item.get("type") == "text":
+                            user_query = item.get("text", "")
+                            break
+                    if user_query:
+                        break
+
+        # Inject memories into system prompt
+        if user_query and self._wrapper._inject_memories:
+            memory_context = self._wrapper._recall_memories(user_query)
+            if memory_context:
+                if system:
+                    kwargs["system"] = f"{system}\n\n{memory_context}"
+                else:
+                    kwargs["system"] = memory_context
+
+        # Make the actual API call
+        response = self._wrapper._client.messages.create(**kwargs)
+
+        # Store conversation
+        if user_query and self._wrapper._store_conversations:
+            if response.content:
+                assistant_output = ""
+                for block in response.content:
+                    if hasattr(block, 'text'):
+                        assistant_output += block.text
+                if assistant_output:
+                    self._wrapper._store_conversation(user_query, assistant_output, model)
+
+        return response
+
+
+def wrap_openai(
+    client: Any,
+    bank_id: str,
+    hindsight_api_url: str = "http://localhost:8888",
+    session_id: Optional[str] = None,
+    store_conversations: bool = True,
+    inject_memories: bool = True,
+    max_memories: Optional[int] = None,
+    recall_budget: str = "mid",
+    verbose: bool = False,
+) -> HindsightOpenAI:
+    """Wrap an OpenAI client with Hindsight memory integration.
+
+    This creates a wrapped client that automatically injects memories
+    and stores conversations when making chat completion calls.
+
+    Args:
+        client: The OpenAI client instance to wrap
+        bank_id: Memory bank ID for memory operations. For multi-user support,
+            use different bank_ids per user (e.g., f"user-{user_id}")
+        hindsight_api_url: URL of the Hindsight API server
+        session_id: Session identifier for conversation grouping
+        store_conversations: Whether to store conversations
+        inject_memories: Whether to inject relevant memories
+        max_memories: Maximum number of memories to inject (None = no limit)
+        recall_budget: Budget level for memory recall (low, mid, high)
+        verbose: Enable verbose logging
+
+    Returns:
+        Wrapped OpenAI client with memory integration
+
+    Example:
+        >>> from openai import OpenAI
+        >>> from hindsight_litellm import wrap_openai
+        >>>
+        >>> client = OpenAI()
+        >>> wrapped = wrap_openai(
+        ...     client,
+        ...     bank_id=f"user-{user_id}",  # Multi-user support via separate banks
+        ... )
+        >>>
+        >>> response = wrapped.chat.completions.create(
+        ...     model="gpt-4",
+        ...     messages=[{"role": "user", "content": "What do you know about me?"}]
+        ... )
+    """
+    return HindsightOpenAI(
+        client=client,
+        bank_id=bank_id,
+        hindsight_api_url=hindsight_api_url,
+        session_id=session_id,
+        store_conversations=store_conversations,
+        inject_memories=inject_memories,
+        max_memories=max_memories,
+        recall_budget=recall_budget,
+        verbose=verbose,
+    )
+
+
+def wrap_anthropic(
+    client: Any,
+    bank_id: str,
+    hindsight_api_url: str = "http://localhost:8888",
+    session_id: Optional[str] = None,
+    store_conversations: bool = True,
+    inject_memories: bool = True,
+    max_memories: Optional[int] = None,
+    recall_budget: str = "mid",
+    verbose: bool = False,
+) -> HindsightAnthropic:
+    """Wrap an Anthropic client with Hindsight memory integration.
+
+    This creates a wrapped client that automatically injects memories
+    and stores conversations when making message calls.
+
+    Args:
+        client: The Anthropic client instance to wrap
+        bank_id: Memory bank ID for memory operations. For multi-user support,
+            use different bank_ids per user (e.g., f"user-{user_id}")
+        hindsight_api_url: URL of the Hindsight API server
+        session_id: Session identifier for conversation grouping
+        store_conversations: Whether to store conversations
+        inject_memories: Whether to inject relevant memories
+        max_memories: Maximum number of memories to inject (None = no limit)
+        recall_budget: Budget level for memory recall (low, mid, high)
+        verbose: Enable verbose logging
+
+    Returns:
+        Wrapped Anthropic client with memory integration
+
+    Example:
+        >>> from anthropic import Anthropic
+        >>> from hindsight_litellm import wrap_anthropic
+        >>>
+        >>> client = Anthropic()
+        >>> wrapped = wrap_anthropic(
+        ...     client,
+        ...     bank_id=f"user-{user_id}",  # Multi-user support via separate banks
+        ... )
+        >>>
+        >>> response = wrapped.messages.create(
+        ...     model="claude-3-5-sonnet-20241022",
+        ...     max_tokens=1024,
+        ...     messages=[{"role": "user", "content": "What do you know about me?"}]
+        ... )
+    """
+    return HindsightAnthropic(
+        client=client,
+        bank_id=bank_id,
+        hindsight_api_url=hindsight_api_url,
+        session_id=session_id,
+        store_conversations=store_conversations,
+        inject_memories=inject_memories,
+        max_memories=max_memories,
+        recall_budget=recall_budget,
+        verbose=verbose,
+    )
diff --git a/hindsight-integrations/litellm/pyproject.toml b/hindsight-integrations/litellm/pyproject.toml
new file mode 100644
index 00000000..b3b4c362
--- /dev/null
+++ b/hindsight-integrations/litellm/pyproject.toml
@@ -0,0 +1,59 @@
+[project]
+name = "hindsight-litellm"
+version = "0.1.0"
+description = "Universal LLM memory integration via LiteLLM - works with 100+ providers"
+readme = "README.md"
+requires-python = ">=3.10"
+license = { text = "MIT" }
+authors = [
+    { name = "Vectorize", email = "support@vectorize.io" }
+]
+keywords = [
+    "ai",
+    "memory",
+    "llm",
+    "litellm",
+    "openai",
+    "anthropic",
+    "groq",
+    "langchain",
+    "agents",
+    "hindsight",
+]
+classifiers = [
+    "Development Status :: 4 - Beta",
+    "Intended Audience :: Developers",
+    "License :: OSI Approved :: MIT License",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+]
+
+dependencies = [
+    "litellm>=1.40.0",
+]
+
+[project.optional-dependencies]
+dev = [
+    "pytest>=7.0.0",
+    "pytest-asyncio>=0.21.0",
+    "pytest-mock>=3.10.0",
+]
+
+[project.urls]
+Homepage = "https://github.com/vectorize-io/hindsight"
+Documentation = "https://github.com/vectorize-io/hindsight/tree/main/hindsight-integrations/litellm"
+Repository = "https://github.com/vectorize-io/hindsight"
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.build.targets.wheel]
+packages = ["hindsight_litellm"]
+
+[tool.pytest.ini_options]
+asyncio_mode = "auto"
+testpaths = ["tests"]
diff --git a/hindsight-integrations/litellm/tests/__init__.py b/hindsight-integrations/litellm/tests/__init__.py
new file mode 100644
index 00000000..abc1ec6c
--- /dev/null
+++ b/hindsight-integrations/litellm/tests/__init__.py
@@ -0,0 +1 @@
+# Tests for hindsight-litellm
diff --git a/hindsight-integrations/litellm/tests/test_integration.py b/hindsight-integrations/litellm/tests/test_integration.py
new file mode 100644
index 00000000..164d4f8a
--- /dev/null
+++ b/hindsight-integrations/litellm/tests/test_integration.py
@@ -0,0 +1,471 @@
+"""Integration tests for hindsight-litellm."""
+
+import pytest
+from unittest.mock import Mock, patch, MagicMock
+from typing import List, Dict, Any
+
+from hindsight_litellm import (
+    configure,
+    enable,
+    disable,
+    is_enabled,
+    cleanup,
+    get_config,
+    is_configured,
+    reset_config,
+    HindsightConfig,
+    MemoryInjectionMode,
+)
+from hindsight_litellm.callbacks import HindsightCallback, get_callback, cleanup_callback
+
+
+class TestConfiguration:
+    """Tests for configuration management."""
+
+    def setup_method(self):
+        """Reset config before each test."""
+        reset_config()
+        disable()
+
+    def teardown_method(self):
+        """Clean up after each test."""
+        cleanup()
+
+    def test_configure_creates_config(self):
+        """Test that configure creates a config object."""
+        config = configure(
+            bank_id="test-agent",
+            hindsight_api_url="http://localhost:8888",
+        )
+
+        assert config is not None
+        assert config.bank_id == "test-agent"
+        assert config.hindsight_api_url == "http://localhost:8888"
+        assert config.enabled is True
+
+    def test_configure_with_all_options(self):
+        """Test configure with all options."""
+        config = configure(
+            hindsight_api_url="http://custom:9999",
+            bank_id="custom-agent",
+            api_key="secret-key",
+            store_conversations=False,
+            inject_memories=False,
+            injection_mode=MemoryInjectionMode.PREPEND_USER,
+            max_memories=5,
+            max_memory_tokens=1000,
+            recall_budget="high",
+            fact_types=["world", "opinion"],
+            document_id="doc-123",
+            enabled=True,
+            excluded_models=["gpt-3.5*"],
+            verbose=True,
+        )
+
+        assert config.hindsight_api_url == "http://custom:9999"
+        assert config.bank_id == "custom-agent"
+        assert config.api_key == "secret-key"
+        assert config.store_conversations is False
+        assert config.inject_memories is False
+        assert config.injection_mode == MemoryInjectionMode.PREPEND_USER
+        assert config.max_memories == 5
+        assert config.max_memory_tokens == 1000
+        assert config.recall_budget == "high"
+        assert config.fact_types == ["world", "opinion"]
+        assert config.document_id == "doc-123"
+        assert config.excluded_models == ["gpt-3.5*"]
+        assert config.verbose is True
+
+    def test_is_configured_without_bank_id(self):
+        """Test is_configured returns False without bank_id."""
+        configure(hindsight_api_url="http://localhost:8888")
+        assert is_configured() is False
+
+    def test_is_configured_with_bank_id(self):
+        """Test is_configured returns True with bank_id."""
+        configure(bank_id="test-agent")
+        assert is_configured() is True
+
+    def test_reset_config(self):
+        """Test reset_config clears the configuration."""
+        configure(bank_id="test-agent")
+        assert is_configured() is True
+
+        reset_config()
+        assert get_config() is None
+        assert is_configured() is False
+
+
+class TestEnableDisable:
+    """Tests for enable/disable functionality."""
+
+    def setup_method(self):
+        """Reset state before each test."""
+        cleanup()
+
+    def teardown_method(self):
+        """Clean up after each test."""
+        cleanup()
+
+    def test_enable_without_config_raises(self):
+        """Test enable raises error without configuration."""
+        with pytest.raises(RuntimeError, match="not configured"):
+            enable()
+
+    def test_enable_registers_callback(self):
+        """Test enable registers callback with LiteLLM."""
+        import litellm
+
+        configure(bank_id="test-agent")
+        enable()
+
+        callback = get_callback()
+        assert callback in litellm.callbacks
+        assert is_enabled() is True
+
+    def test_disable_removes_callback(self):
+        """Test disable removes callback from LiteLLM."""
+        import litellm
+
+        configure(bank_id="test-agent")
+        enable()
+        assert is_enabled() is True
+
+        disable()
+        callback = get_callback()
+        assert callback not in litellm.callbacks
+        assert is_enabled() is False
+
+    def test_enable_idempotent(self):
+        """Test enable is idempotent (can be called multiple times)."""
+        import litellm
+
+        configure(bank_id="test-agent")
+
+        # Enable multiple times
+        enable()
+        enable()
+        enable()
+
+        # Should only have one callback
+        callback = get_callback()
+        assert litellm.callbacks.count(callback) == 1
+
+
+class TestCallback:
+    """Tests for the HindsightCallback class."""
+
+    def setup_method(self):
+        """Reset state before each test."""
+        cleanup()
+
+    def teardown_method(self):
+        """Clean up after each test."""
+        cleanup()
+
+    def test_extract_user_query_simple(self):
+        """Test extracting user query from simple messages."""
+        callback = HindsightCallback()
+        messages = [
+            {"role": "system", "content": "You are helpful."},
+            {"role": "user", "content": "What is the capital of France?"},
+        ]
+
+        query = callback._extract_user_query(messages)
+        assert query == "What is the capital of France?"
+
+    def test_extract_user_query_from_last_user_message(self):
+        """Test extracting query from last user message."""
+        callback = HindsightCallback()
+        messages = [
+            {"role": "user", "content": "First question"},
+            {"role": "assistant", "content": "First answer"},
+            {"role": "user", "content": "Second question"},
+        ]
+
+        query = callback._extract_user_query(messages)
+        assert query == "Second question"
+
+    def test_extract_user_query_structured_content(self):
+        """Test extracting query from structured content (vision)."""
+        callback = HindsightCallback()
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What's in this image?"},
+                    {"type": "image_url", "image_url": {"url": "http://example.com/img.png"}},
+                ],
+            },
+        ]
+
+        query = callback._extract_user_query(messages)
+        assert query == "What's in this image?"
+
+    def test_extract_user_query_multiple_text_parts(self):
+        """Test extracting query with multiple text parts."""
+        callback = HindsightCallback()
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "First part."},
+                    {"type": "text", "text": "Second part."},
+                ],
+            },
+        ]
+
+        query = callback._extract_user_query(messages)
+        assert query == "First part. Second part."
+
+    def test_format_memories(self):
+        """Test formatting memories into context string."""
+        callback = HindsightCallback()
+        config = HindsightConfig(bank_id="test", max_memories=10, verbose=False)
+
+        memories = [
+            {"text": "User likes Python", "fact_type": "world", "weight": 0.95},
+            {"text": "User works at Google", "fact_type": "world", "weight": 0.8},
+        ]
+
+        formatted = callback._format_memories(memories, config)
+
+        assert "Relevant Memories" in formatted
+        assert "User likes Python" in formatted
+        assert "User works at Google" in formatted
+        assert "[WORLD]" in formatted
+
+    def test_format_memories_with_verbose(self):
+        """Test formatting memories with verbose mode shows weights."""
+        callback = HindsightCallback()
+        config = HindsightConfig(bank_id="test", max_memories=10, verbose=True)
+
+        memories = [
+            {"text": "User likes Python", "fact_type": "world", "weight": 0.95},
+        ]
+
+        formatted = callback._format_memories(memories, config)
+
+        assert "relevance: 0.95" in formatted
+
+    def test_inject_memories_as_system_message(self):
+        """Test injecting memories as system message."""
+        callback = HindsightCallback()
+        config = HindsightConfig(
+            bank_id="test",
+            injection_mode=MemoryInjectionMode.SYSTEM_MESSAGE,
+        )
+
+        messages = [
+            {"role": "user", "content": "Hello"},
+        ]
+        memory_context = "# Relevant Memories\n1. User is John"
+
+        result = callback._inject_memories_into_messages(messages, memory_context, config)
+
+        assert len(result) == 2
+        assert result[0]["role"] == "system"
+        assert "Relevant Memories" in result[0]["content"]
+        assert result[1]["role"] == "user"
+
+    def test_inject_memories_prepend_to_existing_system(self):
+        """Test injecting memories appends to existing system message."""
+        callback = HindsightCallback()
+        config = HindsightConfig(
+            bank_id="test",
+            injection_mode=MemoryInjectionMode.SYSTEM_MESSAGE,
+        )
+
+        messages = [
+            {"role": "system", "content": "You are helpful."},
+            {"role": "user", "content": "Hello"},
+        ]
+        memory_context = "# Relevant Memories\n1. User is John"
+
+        result = callback._inject_memories_into_messages(messages, memory_context, config)
+
+        assert len(result) == 2
+        assert result[0]["role"] == "system"
+        assert "You are helpful." in result[0]["content"]
+        assert "Relevant Memories" in result[0]["content"]
+
+    def test_inject_memories_prepend_user_mode(self):
+        """Test injecting memories in prepend_user mode."""
+        callback = HindsightCallback()
+        config = HindsightConfig(
+            bank_id="test",
+            injection_mode=MemoryInjectionMode.PREPEND_USER,
+        )
+
+        messages = [
+            {"role": "user", "content": "What's my name?"},
+        ]
+        memory_context = "# Relevant Memories\n1. User is John"
+
+        result = callback._inject_memories_into_messages(messages, memory_context, config)
+
+        assert len(result) == 1
+        assert result[0]["role"] == "user"
+        assert "Relevant Memories" in result[0]["content"]
+        assert "What's my name?" in result[0]["content"]
+
+    def test_should_skip_model_exact_match(self):
+        """Test model exclusion with exact match."""
+        callback = HindsightCallback()
+        config = HindsightConfig(
+            bank_id="test",
+            excluded_models=["gpt-3.5-turbo"],
+        )
+
+        assert callback._should_skip_model("gpt-3.5-turbo", config) is True
+        assert callback._should_skip_model("gpt-4", config) is False
+
+    def test_should_skip_model_wildcard(self):
+        """Test model exclusion with wildcard pattern."""
+        callback = HindsightCallback()
+        config = HindsightConfig(
+            bank_id="test",
+            excluded_models=["gpt-3.5*", "claude-instant-*"],
+        )
+
+        assert callback._should_skip_model("gpt-3.5-turbo", config) is True
+        assert callback._should_skip_model("gpt-3.5-turbo-16k", config) is True
+        assert callback._should_skip_model("claude-instant-1.2", config) is True
+        assert callback._should_skip_model("gpt-4", config) is False
+        assert callback._should_skip_model("claude-3-opus", config) is False
+
+
+class TestDeduplication:
+    """Tests for conversation deduplication."""
+
+    def setup_method(self):
+        """Reset state before each test."""
+        cleanup()
+
+    def teardown_method(self):
+        """Clean up after each test."""
+        cleanup()
+
+    def test_compute_conversation_hash(self):
+        """Test computing conversation hash."""
+        callback = HindsightCallback()
+
+        hash1 = callback._compute_conversation_hash("Hello", "Hi there!")
+        hash2 = callback._compute_conversation_hash("Hello", "Hi there!")
+        hash3 = callback._compute_conversation_hash("Hello", "Different response")
+
+        # Same content should produce same hash
+        assert hash1 == hash2
+        # Different content should produce different hash
+        assert hash1 != hash3
+
+    def test_compute_conversation_hash_case_insensitive(self):
+        """Test that hash is case insensitive."""
+        callback = HindsightCallback()
+
+        hash1 = callback._compute_conversation_hash("HELLO", "HI THERE!")
+        hash2 = callback._compute_conversation_hash("hello", "hi there!")
+
+        assert hash1 == hash2
+
+    def test_is_duplicate_first_time(self):
+        """Test first occurrence is not a duplicate."""
+        callback = HindsightCallback()
+
+        result = callback._is_duplicate("abc123")
+
+        assert result is False
+
+    def test_is_duplicate_second_time(self):
+        """Test second occurrence is a duplicate."""
+        callback = HindsightCallback()
+
+        callback._is_duplicate("abc123")  # First time
+        result = callback._is_duplicate("abc123")  # Second time
+
+        assert result is True
+
+    def test_is_duplicate_different_hashes(self):
+        """Test different hashes are not duplicates."""
+        callback = HindsightCallback()
+
+        callback._is_duplicate("abc123")
+        result = callback._is_duplicate("xyz789")
+
+        assert result is False
+
+
+class TestContextManager:
+    """Tests for the hindsight_memory context manager."""
+
+    def setup_method(self):
+        """Reset state before each test."""
+        cleanup()
+
+    def teardown_method(self):
+        """Clean up after each test."""
+        cleanup()
+
+    def test_context_manager_enables_and_disables(self):
+        """Test context manager enables and disables correctly."""
+        from hindsight_litellm import hindsight_memory
+
+        assert is_enabled() is False
+
+        with hindsight_memory(bank_id="test-agent"):
+            assert is_enabled() is True
+            assert get_config().bank_id == "test-agent"
+
+        assert is_enabled() is False
+
+    def test_context_manager_restores_previous_config(self):
+        """Test context manager restores previous configuration."""
+        from hindsight_litellm import hindsight_memory
+
+        # Set up initial config
+        configure(bank_id="original-agent")
+        enable()
+        assert get_config().bank_id == "original-agent"
+
+        # Use context manager with different config
+        with hindsight_memory(bank_id="temporary-agent"):
+            assert get_config().bank_id == "temporary-agent"
+
+        # Should restore original config
+        assert get_config().bank_id == "original-agent"
+        assert is_enabled() is True
+
+    def test_context_manager_with_fact_types(self):
+        """Test context manager with fact_types parameter."""
+        from hindsight_litellm import hindsight_memory
+
+        with hindsight_memory(bank_id="test-agent", fact_types=["world", "opinion"]):
+            config = get_config()
+            assert config.fact_types == ["world", "opinion"]
+
+
+class TestFactTypes:
+    """Tests for fact_types configuration."""
+
+    def setup_method(self):
+        """Reset config before each test."""
+        reset_config()
+
+    def teardown_method(self):
+        """Clean up after each test."""
+        cleanup()
+
+    def test_configure_with_fact_types(self):
+        """Test configuring with fact_types."""
+        config = configure(
+            bank_id="test-agent",
+            fact_types=["world", "agent", "opinion"],
+        )
+
+        assert config.fact_types == ["world", "agent", "opinion"]
+
+    def test_configure_without_fact_types(self):
+        """Test configuring without fact_types defaults to None."""
+        config = configure(bank_id="test-agent")
+
+        assert config.fact_types is None