* fix(llamaindex): use uuid for document_id and sync version metadata
- Replace timestamp-based document_id with uuid4 hex to prevent
collisions on rapid retains (timestamp_ms can duplicate in tight loops)
- Sync __version__ in __init__.py to match pyproject.toml (0.1.2)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(docs): pass memory to run() instead of ReActAgent constructor
LlamaIndex 0.14.x ReActAgent does not accept a memory parameter in
its constructor — it's silently dropped via **kwargs. Memory must be
passed to agent.run(memory=...) where AgentWorkflow picks it up.
Also fixes the undefined `tools` variable (now `tools=[]`).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(llamaindex): strip ReAct reasoning traces from retained assistant messages
HindsightMemory.put/aput now extracts only the final Answer: text from
assistant messages containing ReAct reasoning (Thought:/Action:/Observation:
prefixes), preventing internal reasoning traces from polluting long-term memory.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(llamaindex): fix docstring example to pass memory to run()
The HindsightMemory class docstring showed the broken pattern of passing
memory= to the ReActAgent constructor, which silently drops it. Updated
to show the correct pattern: pass memory to agent.run().
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Document the new configurable base URL for the ZeroEntropy reranker
provider added in #766. Also fix a type error where RERANK_URL was
renamed to rerank_url but one usage was missed.
The automatic memory example referenced an undefined `tools` variable.
Since HindsightMemory handles retain/recall transparently, no tools
are needed — use an empty list.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Fixes#771 — two trailing commas in openclaw.plugin.json caused OpenClaw's
strict JSON parser to reject the plugin manifest during installation.
Also adds JSON validation tests for both the openclaw plugin manifest and
the claude-code hooks.json so CI catches invalid JSON before release.
* refactor(llamaindex): merge two packages into single hindsight-llamaindex
Merge `llama-index-tools-hindsight` and `llama-index-memory-hindsight` into
a single `hindsight-llamaindex` package following our naming convention.
- Rename package to `hindsight-llamaindex` (Python module: `hindsight_llamaindex`)
- Move HindsightToolSpec and HindsightMemory into the same package
- Delete `llamaindex-memory/` directory
- Add CI test job for llamaindex integration
- Update docs, blog post, and integrations.json
* fix(blog): update llamaindex blog post for merged package
- Move date to 2026-03-30
- Add HindsightMemory (automatic BaseMemory) pattern
- Fix "bank must exist first" pitfall — mission auto-creates
- Align all code examples with docs page
- Update architecture diagram to show both patterns
* fix(docs): add llamaindex/openai icons, rename Codex
- Add llamaindex.png and openai.png icons
- Rename "OpenAI Codex CLI" to "Codex" in integrations.json and docs
- Use openai.png icon for Codex integration
* feat(api): add duration_ms to audit log entries
Server-computed duration in milliseconds (started_at → ended_at) on
the list audit logs endpoint. Null when ended_at is not set.
Closes#749
* feat(api): add duration_ms to audit log entries and type audit endpoints
- Add server-computed duration_ms (started_at → ended_at) to audit log
list response. Null when ended_at is not set.
- Add typed Pydantic response models for both audit log endpoints
(list and stats) so they appear in the OpenAPI spec.
- Regenerate OpenAPI spec and all client SDKs.
Closes#749
* chore: regenerate docs skill after audit log response models
* fix(mcp): handle Claude Code GET probe and make stateless_http configurable (#751)
Claude Code v2.1.84+ sends a GET to /mcp/ before POST initialize,
which fails with 405 (stateless) or 400 (stateful). Intercept
sessionless GET requests in MCPMiddleware and return 200 OK so the
client proceeds to POST initialize.
Also make stateless_http configurable via HINDSIGHT_API_MCP_STATELESS
(default: false/stateful) instead of hardcoding true.
Closes#751
* docs: add HINDSIGHT_API_MCP_STATELESS to configuration reference
* Convert codex tool_choice test to pytest style
Follow-up to #734: replace unittest.TestCase + manual sys.path
manipulation with idiomatic pytest + @pytest.mark.asyncio,
matching the rest of the test suite.
* Fix test_hierarchical_fields_categorization for new configurable fields
Update expected count from 20 to 21 and add assertions for fields
added by recent PRs: retain_default_strategy, retain_strategies,
max_observations_per_scope, reflect_source_facts_max_tokens,
llm_gemini_safety_settings, mcp_enabled_tools.
* Add LlamaIndex doc to v0.4 versioned docs and sidebars
The LlamaIndex integration doc was added to docs/ (next version) in
#672 but not to versioned_docs/version-0.4/, causing a broken link
on the /integrations page which resolves to the latest version.
* Regenerate docs skill references
Run generate-docs-skill.sh to pick up new integration pages
(codex, llamaindex) and updated configuration docs.
* Add Codex integration doc to v0.4 versioned docs and sidebar
Same issue as LlamaIndex: doc was added to docs/ (next) but not
versioned_docs/version-0.4/, causing broken link on /integrations.
create_bank_hnsw_indexes() hardcoded USING hnsw regardless of the configured
vector extension, causing "column cannot have more than 2000 dimensions for
hnsw index" when using pgvectorscale or vchord with high-dimensional embeddings.
Now reads get_config().vector_extension and uses the appropriate index type:
- pgvector → USING hnsw
- pgvectorscale → USING diskann
- vchord → USING vchordrq
Closes#738
Verbose mode was the only extraction mode that skipped injecting the
retain_mission FOCUS section into its prompt template. Users who set a
retain_mission got no filtering when using verbose mode.
* fix(codex): cleanup dead code and add to release lifecycle
- Remove orphaned reflect() method from client.py (leftover from dropped auto-mode)
- Remove dead retainToolCalls config default (never wired through)
- Add codex to release-integration.sh valid integrations
- Add settings.json version fallback to release script
- Add codex CI test job in test.yml
- Add codex to integrations.json registry
* docs(codex): add changelog page and link from integration docs
* feat(codex): add hosted installer script (get-codex)
Add self-contained installer at hindsight.vectorize.io/get-codex that
downloads scripts from GitHub, configures hooks, and supports local/cloud
mode selection — no git clone required.
Update docs and README to use the one-liner install:
curl -fsSL https://hindsight.vectorize.io/get-codex | bash
* chore(codex): remove install.sh in favor of hosted get-codex
* fix(docs): use /next/ prefix for codex changelog link
* fix(docs): use GitHub link for codex changelog back-link
* feat(codex): add Hindsight memory integration for OpenAI Codex CLI
Hooks-based integration that gives Codex CLI long-term memory via Hindsight.
Three hooks keep memory in sync: SessionStart (daemon pre-warm), UserPromptSubmit
(recall + context injection), Stop (retain conversation to memory).
Key differences from the Claude Code integration:
- Codex transcript format: JSONL with {msg: {type, message}} (user_message/agent_message)
- No CODEX_PLUGIN_ROOT env var — install.sh writes hooks.json with absolute paths
- State stored in ~/.hindsight/codex/state/ (not CLAUDE_PLUGIN_DATA)
- No async: true in hooks (not supported by Codex)
- No SessionEnd event
- hooks.json written to ~/.codex/hooks.json with codex_hooks = true in config.toml
* fix(codex): fix transcript parser for actual Codex disk format
Codex stores sessions as rollout-*.jsonl with response_item entries:
User: {type:response_item, payload:{type:message, role:user, content:[{type:input_text, text:...}]}}
Assistant: {type:response_item, payload:{type:message, role:assistant, phase:final_answer, content:[{type:output_text, text:...}]}}
Previous parser expected an undocumented {msg:{type:user_message}} format from the Rust protocol spec
that does not match the actual on-disk storage format.
* feat(codex): add reflect mode to UserPromptSubmit hook
Add recallMode config option (default: 'recall') that switches the
UserPromptSubmit hook between:
- 'recall': existing behavior, fast raw facts list
- 'reflect': agentic synthesis loop, returns coherent prose answer
Also adds reflect() method to HindsightClient and HINDSIGHT_RECALL_MODE
env var override. Reflect uses a 25s timeout (vs 10s for recall).
* feat(codex): auto mode for recall/reflect selection
Add recallMode: 'auto' (new default) that picks the operation per-query:
- Synthesis patterns (what do you know, what's my, summarize, etc.) → reflect
- All other prompts → recall (fast, raw facts, better for code tasks)
* feat(codex): add automated test suite and finalize recall-only mode
* docs(codex): add docs page and sidebar entry for Codex CLI integration
* fix(hermes): convert lifecycle hooks to sync for hermes-agent 0.5.0 compatibility
hermes-agent 0.5.0 calls plugin hooks synchronously via invoke_hook(),
but our pre_llm_call/post_llm_call were async — coroutines were never
awaited, so recall context injection and auto-retain silently did nothing.
Switch hooks to sync client methods and add integration tests using
the real hermes-agent PluginManager.
* fix(hermes): use proper hermes-agent dep with uv source override
Replace inline git URL with standard `hermes-agent>=0.5.0` version
constraint plus `[tool.uv.sources]` to resolve from the git tag until
0.5.0 lands on PyPI.
* feat: add LlamaIndex integration for Hindsight
Add hindsight-llamaindex package providing persistent memory tools for
LlamaIndex agents via the native BaseToolSpec pattern. Includes retain,
recall, and reflect tools, a convenience factory, global config, full
test suite, docs page, blog post, and integrations.json entry.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: address PR review feedback for llamaindex integration
- Fix ReActAgent API: from_tools() → constructor, chat() → await run()
- Add create_bank step to all quickstart examples
- Add production patterns section to docs (tags, error handling, bank lifecycle)
- Add memory scoping recommendation to README
- Add when-not-to-use section to blog post
- Add LlamaIndex compatibility tests (agent acceptance, FunctionTool.call)
- Fix self-hosted auth wording in cookbook notebook
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: use async client methods and asyncio.run() for runnable examples
- Use await client.acreate_bank() instead of sync create_bank() to
avoid "event loop already running" errors in notebooks and async contexts
- Wrap plain Python examples in async def main() + asyncio.run(main())
so they are copy-paste runnable as scripts
- Add Jupyter notebook tip to docs showing top-level await pattern
- Bank lifecycle example in docs now uses async acreate_bank
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: add async tool methods to avoid event loop conflicts
HindsightToolSpec now provides both sync and async tool implementations
using LlamaIndex's (sync_fn, async_fn) tuple pattern in spec_functions.
Async agents (ReActAgent, etc.) use aretain/arecall/areflect natively,
avoiding the "Timeout context manager should be used inside a task"
error that occurred when sync _run_async() was called from within an
active event loop.
- Add aretain_memory, arecall_memory, areflect_on_memory async methods
- Extract shared kwargs builders (_retain_kwargs, _recall_kwargs, etc.)
- spec_functions now uses tuples: [("retain_memory", "aretain_memory"), ...]
- Tests verify tools have both sync fn and async fn set
- Notebook verified end-to-end with nbclient against local Hindsight
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* chore: remove blog post from integration PR
The blog post will be pulled in separately from its own PR.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Address PR review: add context label, document_id auto-gen, bank mission, graceful errors
- Add `retain_context` param (default: "llamaindex") as source label on retain ops
- Auto-generate `document_id` as `{session_id}-{timestamp_ms}` when not provided
- Add `retain_async` param (default: True) for non-blocking retain processing
- Add `mission` param for automatic bank creation/management on first use
- Change error handling from raising HindsightError to graceful log + return message
- Add per-operation timeout constants in _client.py
- Add `context` and `mission` fields to config.py and configure()
- Update docs: document as standalone package (not LlamaHub), new params, patterns
- Tests: 51 passing (up from 34), covering all new features
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* Restructure to LlamaIndex namespace packages + add BaseMemory implementation
Tools package (llama-index-tools-hindsight):
- Restructured from hindsight_llamaindex/ to llama_index/tools/hindsight/
- Import: from llama_index.tools.hindsight import HindsightToolSpec
- Follows PEP 420 implicit namespace package convention
- Removed retain_async param (client.retain() doesn't support async_processing)
Memory package (llama-index-memory-hindsight):
- New package: llama_index/memory/hindsight/
- HindsightMemory(BaseMemory) for automatic memory
- put() auto-retains user/assistant messages to Hindsight
- get(input) auto-recalls relevant memories, prepends as system message
- Graceful error handling, bank mission management, document_id generation
- 28 unit tests passing
Both packages follow LlamaIndex community conventions for future LlamaHub submission.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
A 429 usage_limit_reached response during verify_connection() caused the
server to refuse to start entirely. Quota exhaustion is not a configuration
error — the server should start and serve retain/recall requests normally,
it just can't make LLM calls until the quota resets.
Co-authored-by: Marco Rutsch <marco@rutimka.de>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(openclaw): configurable logging with structured output
Replace raw console.log/warn/error spam with a structured logger.
New plugin settings: logLevel, logSummaryIntervalMs, logCompact.
Bank mission log demoted to verbose-only. Retain/recall batched
into periodic summaries. Each recall now shows memory count injected.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* use api.logger for framework-consistent output, show autoRecall/autoRetain on init
Route all log output through OpenClaw's api.logger instead of raw console
calls. Matches mem0 plugin style. Startup now shows mode + feature flags.
Dropped logCompact setting (framework handles formatting). Added subtle
slate-blue color to hindsight prefix for visual differentiation.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* add bank name to init and summary logs, fix singular/plural consistency
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* rename log levels to standard: off, error, warning, info, debug
Per review feedback — use standard level names instead of custom ones.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: billy <billy@oclaw.local>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: Nicolò Boschi <boschi1997@gmail.com>
Add optional filter_mcp_tools() method to OperationValidatorExtension.
Called during tools/list after bank-level mcp_enabled_tools filtering.
Extensions can override to hide MCP tools per-user-per-bank based on
access policies. Default returns all tools unchanged.
- Add filter_mcp_tools to OperationValidatorExtension with default pass-through
- Wire into _get_enabled_tools in _apply_bank_tool_filtering
- Move _ALL_TOOLS to mcp_tools.py to avoid circular import (re-exported from mcp.py)
- Fail-open: if filter raises, log warning and return unfiltered tools
- Enforce ceiling: validator can narrow but never expand beyond bank config
- Add 8 tests: default, filtering, empty set, integration, composition,
can't-add-tools, exception fail-open, no-validator passthrough
* fix: parse query params from base_url in OpenAI embeddings client
The OpenAI-compatible LLM provider already parses query parameters
(e.g. ?api-version=xxx for Azure OpenAI) from the base_url and passes
them as default_query to the OpenAI client. However, the OpenAI
embeddings provider did not do this, causing Azure OpenAI embeddings
to fail with 404 errors at runtime.
This applies the same URL parsing logic from the LLM provider to the
embeddings provider, enabling Azure OpenAI embeddings to work correctly.
* ci: add workflow to build fork Docker image
* ci: add slim image build (no local models)
* ci: remove fork build workflow per review request
---------
Co-authored-by: Antoine Khater <ak@ptgroup.eu>
* fix(claude-code): implement tool_choice support for forced tool calls
The call_with_tools() method now properly handles the tool_choice parameter
to force specific tool calls. Previously, the parameter was accepted but ignored,
causing the reflect agent to fail when trying to force specific tools on each
iteration.
Fixes#732
Changes:
- When tool_choice forces a specific function: filter allowed_tools to only
that tool (with mcp prefix) and add a strong system prompt instruction
- When tool_choice is 'required': add instruction that model must call at
least one tool
- When tool_choice is 'none': clear allowed_tools and mcp_servers to disable
all tools
- When tool_choice is 'auto' (default): no change (existing behavior)
This matches the approach used in the OpenAI provider while adapting to the
Claude Agent SDK's lack of native tool_choice parameter by using allowed_tools
filtering and system prompt instructions.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* style: fix ruff formatting in alembic migration
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat: add max_observations_per_scope bank config
Adds a configurable limit on the number of observations per tag scope.
When the limit is reached, consolidation only updates/deletes existing
observations — no new ones are created. Enforcement is done via a
constrained Pydantic response model (max_length on creates list) so the
LLM structurally cannot exceed the limit, plus prompt guidance.
- Config: HINDSIGHT_API_MAX_OBSERVATIONS_PER_SCOPE (-1 = unlimited)
- Reorder action execution: deletes → updates → creates
- Dynamic _ConsolidationBatchResponse with max_length constraint
- Prompt CAPACITY CONSTRAINT section when near/at limit
- Observations with no tags skip the limit entirely
- Control plane UI field + docs
* fix: strengthen max_observations tests with mock LLM + defensive truncation
- Rewrite integration tests to use MockLLM with deterministic responses
(one observation per fact) instead of relying on real LLM behavior
- Add defensive truncation in _consolidate_batch_with_llm as belt-and-
suspenders — catches LLM providers that ignore JSON schema max_length
- Tests now assert exact counts, not just upper bounds
The auto-recall timeout was hardcoded to 10s but recall with budget=high
can take 13s+. This adds a configurable recallTimeoutMs option (default:
10000ms) so users can increase the timeout when using higher recall budgets.
Also adds recallInjectionPosition to the plugin schema (it was already
implemented in code but missing from the JSON schema validation, causing
config rejection).
Co-authored-by: Marco Rutsch <marco@rutimka.de>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Replace start_ui()/stop_ui()/is_ui_running() methods with declarative
constructor flags (ui, ui_port, ui_hostname). UI lifecycle now follows
the daemon automatically - starts in _ensure_started, stops in _cleanup.
Add integration test verifying UI starts and can reach the dataplane
via the control plane's /api/health endpoint. Add Node.js setup to
test-hindsight-all CI job to support the UI test.
* Add blog: How We Built a 4-Way Parallel Hybrid Search System
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Add cover image for parallel hybrid search post
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Update parallel hybrid search post date to 2026-03-27
* Set author to chrislatimer
* Update recall docs link
* review: align blog post to actual retrieval code
- Reframe as evolutionary narrative (V1 asyncio.gather → connection sharing)
- Add missing reranker section (cross-encoder + multiplicative boost scoring)
- Replace MPFP references with LinkExpansion (3-signal CTE)
- Fix SQL to match actual UNION ALL approach, explain CTE planner issue
- Fix acquire_with_retry, index types (ivfflat→HNSW), fusion code
- Remove fabricated perf numbers
- Add alpha calibration rationale and connection contention insight
* add nicoloboschi and benfrank241 as co-authors
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-authored-by: Nicolò Boschi <boschi1997@gmail.com>
The stats endpoint JOINs memory_links to memory_units just to filter
by bank_id. With 8.2M+ links per bank this takes 18+ seconds, and
the control plane polls every 10s — perpetually blocking the server.
Add bank_id column directly to memory_links so the query can filter
on ml.bank_id instead of mu.bank_id, letting Postgres push the filter
down before the JOIN.
- Migration: add bank_id TEXT NOT NULL, backfill from memory_units
- All 4 INSERT paths (temporal, semantic, entity, causal) now write bank_id
- Stats query filters on ml.bank_id instead of mu.bank_id
* fix(migrations): use HINDSIGHT_API_MIGRATION_DATABASE_URL when set
Session-level advisory locks are broken when the database URL goes
through PgBouncer in transaction mode: the backend connection is
returned to the pool on COMMIT, orphaning the lock, so multiple pods
can simultaneously run migrations for the same schema.
When HINDSIGHT_API_MIGRATION_DATABASE_URL is set, use it for both
the advisory lock connection and the Alembic run. Callers should
point this at the direct PostgreSQL endpoint (bypassing the pooler)
so the session-level lock is held for the full migration duration.
* refactor(migrations): move MIGRATION_DATABASE_URL to standard config
Wire HINDSIGHT_API_MIGRATION_DATABASE_URL through HindsightConfig
instead of reading os.getenv() directly in migrations.py. Add the
field to the dataclass, from_env(), log_config(), all call sites,
.env.example, and the configuration docs page.
* fix: update test mocks for migration_database_url kwarg and regenerate docs skill
---------
Co-authored-by: Nicolò Boschi <boschi1997@gmail.com>
* fix: silence noisy google_genai.models INFO logging
The google-genai SDK logs "AFC is enabled with max remote calls: 10"
at INFO level on every initialization. Set its logger to WARNING.
* fix: regenerate docs skill in release-integration script
The release script generates changelog/SDK pages but never re-ran
generate-docs-skill.sh, causing CI to fail with out-of-sync skill
files after every integration release. Now it regenerates the skill
and includes the output in the release commit.
Also adds the missing ag2 skill files from the latest release.
* fix(migration): use IF EXISTS when dropping chunk FK constraint
The migration unconditionally dropped memory_units_chunk_fkey, but
depending on the order in which migrations were applied the constraint
may not exist. Use raw SQL with IF EXISTS so the drop is safe regardless.
* fix(migration): make chunk FK add idempotent with DO block
The previous fix only handled the DROP side with IF EXISTS. The ADD side
could still fail with DuplicateObject when the FK already existed on a
schema that was provisioned after the base migration ran.
Wrap the ADD CONSTRAINT in a DO block to catch duplicate_object and
continue, making the migration fully idempotent in both directions.
Port fixes from #461 (claude_code_llm) to codex_llm:
- Replace json.dumps(result) with result.model_dump_json() for Pydantic models to fix TypeError during consolidation
- Wrap record_llm_call tracing block in try/except so logging failures never propagate to retry handler
Co-authored-by: Marco Rutsch <marco@rutimka.de>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
