The warning that a merge on GitHub gets overwritten was living in the
repository description. That is the wrong surface for it twice over: the
description is what every catalogue and topic shelf renders as our card,
so ~120 of its characters were spent on an operational note aimed at
contributors, and contributors do not read the description — they read
the README before opening a pull request.
Banner sits directly under the tagline: canonical repository, why it is
self-hosted, what happens to a merge performed here, and that issues,
stars and pull requests are still welcome on this side.
The repository, the npm package and the container image are now
fleet-memory; RCLL stays the product and documentation brand (rcll.ai).
- server.json / package.json: io.github.holetron-lab/fleet-memory,
npm identifier fleet-memory-mcp
- env chain is two names, not three: FLEET_URL/FLEET_BANK with the
pre-rebrand HINDSIGHT_URL/MEMPALACE_BANK still read as a fallback.
RCLL_URL/RCLL_BANK and bank rcll-main never shipped — nothing was
published under rcll-mcp — so they are dropped rather than carried.
- default bank is fleet-main; hindsight-mempalace-mcp@1.0.0 users are
told on stderr that their memory is in mempalace-main
- rooms are described as topic scoping over one shared store, not
per-agent isolation: the read path filters on the room list the
caller passes and has no notion of caller identity
Three things found by actually running the server over stdio against a live backend
instead of describing it.
1. recall ignored 'limit'. The backend treats it as a retrieval hint and returns
everything inside its own token budget — 113 facts, 43 KB of JSON, for a request
that asked for 2. The tool's own description ('Max results') was therefore false,
and every recall spent roughly 10k tokens of the caller's context. The MCP layer
now enforces the requested limit and reports 'N of M' when it truncated:
limit=2 goes from 43541 to 1000 chars.
2. serverInfo.version said 1.0.0 while package.json and server.json say 0.1.0. The
MCP registry validates the resolved package against the manifest, so a version
that disagrees with the tarball is a review failure.
3. The README claimed 'Claude Code, OpenClaw, Cursor, etc.' None of those was tested.
Replaced with what was measured — MCP 2025-06-18, initialize/tools/list/tools/call
round-trip, memory_recall returning real results over stdio — plus the Claude Code
config we actually run.
The default memory bank becomes 'rcll-main'. hindsight-mempalace-mcp@1.0.0 (published,
~145 downloads/month) defaulted to 'mempalace-main', so an install that never set the
variable and switches packages would open a different, empty bank. MEMPALACE_BANK is
still honoured, and when neither variable is set the server now prints which bank it
picked and how to keep reading the old one, instead of choosing silently.
rcll-split.py moves to scripts/ — it documents how the fork series was built, which
is not a root-level artifact of the product.
* fix: add PEP 561 py.typed marker to all Python packages
Add empty py.typed marker files to all 13 Python packages that were
missing them. Only hindsight-integrations/autogen already had one.
Per PEP 561, packages that wish to support type checking must include
a py.typed marker file. Without it, type checkers (mypy, pyright) treat
the package as untyped and skip all inline type annotations.
Fixes#965
* fix: ensure py.typed markers survive client regeneration
Add touch commands in generate-clients.sh to recreate PEP 561 py.typed
marker files after the OpenAPI generator runs, since the script deletes
and regenerates the hindsight_client_api directory.
---------
Co-authored-by: r266-tech <r266-tech@users.noreply.github.com>
Reflect's tool functions (tool_search_observations, tool_recall) call
recall_async with the user's original request_context, which has
internal=False. The usage metering extension sees these as user-facing
recall operations and bills them separately — double-charging the
customer for recalls that are already included in the reflect operation
cost.
Fix: wrap request_context with dataclasses.replace(internal=True) before
passing to recall_async. This matches the pattern used by consolidation,
which already creates an internal RequestContext for its sub-operations.
The internal flag causes the metering extension to:
- Record the usage as "internal_recall" (tracked but not billed)
- Skip credit deduction entirely
Observed impact: a single reflect call was generating 2 extra billed
recall entries (one from tool_search_observations, one from tool_recall),
inflating the customer's recall token count by ~26 tokens per reflect.
Adds an OAuth 2.1 proxy Worker that connects cloud MCP clients
(claude.ai, Claude Code, Codex) to a self-hosted Hindsight instance
via Cloudflare Workers and Tunnel.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
compute_semantic_links_ann created a TEMP TABLE outside any transaction,
then ran a TRUNCATE / COPY / SELECT / DROP sequence as separate statements
on the same asyncpg connection. This is fine against a direct Postgres
connection but fails intermittently when the caller is routed through
PgBouncer in transaction pool mode:
CREATE TEMP TABLE IF NOT EXISTS _ann_seeds (...) -- backend A
TRUNCATE _ann_seeds -- backend B -> FAILS
Temp tables are session-scoped to the backend that created them. In
PgBouncer transaction mode the backend is only pinned to the client for
the duration of an actual transaction, so between standalone statements
the pooler can (and under concurrency, will) rebind the client to a
different backend. When that happens the _ann_seeds table disappears
and the follow-up statement fails with:
relation "_ann_seeds" does not exist
Symptom: ~3% of sync retain calls (2 of 61) failed the Hindsight Cloud
smoke test on a recent hindsight-dev deploy. Async retains are masked
by the 3-attempt retry loop so they usually eventually succeed.
Fix: wrap the CREATE TEMP TABLE -> COPY -> SELECT sequence in a single
`async with conn.transaction():` block, and use ON COMMIT DROP so the
temp table is transaction-scoped and auto-cleaned at commit. Also
switch `SET hnsw.ef_search = 60` to `SET LOCAL` so the tuning is
transaction-scoped and no longer leaks onto the pooled backend for
subsequent recall queries. Drop the now-unnecessary manual TRUNCATE,
explicit DROP TABLE, and RESET hnsw.ef_search.
The function docstring still correctly describes this as running on a
separate connection outside the surrounding write transaction — this
change only adds an inner transaction around the ANN work itself to
keep the temp table visible to PgBouncer.
Tests:
- Add TestComputeSemanticLinksAnnPgBouncerSafety with 5 regression
tests using a mocked connection. These are structural asserts — they
check that the function enters conn.transaction(), uses ON COMMIT DROP,
uses SET LOCAL, and does not reintroduce manual TRUNCATE / DROP /
RESET calls. They would have caught the original bug if they had
existed, and will catch any future reversion.
* refactor(openclaw)!: read config from plugin config instead of process.env
The plugin loaded credentials and runtime settings from environment
variables (HINDSIGHT_API_LLM_*, HINDSIGHT_EMBED_API_*, HINDSIGHT_BANK_ID)
plus auto-detection of OPENAI_API_KEY / ANTHROPIC_API_KEY / GEMINI_API_KEY
/ GROQ_API_KEY. That tripped OpenClaw's install-scanner env-harvesting
rule and bypassed the framework's first-class SecretRef resolution.
Switch to reading from the plugin config exclusively, with secrets
configured via 'openclaw config set ... --ref-source env|file|exec'.
Combined with the daemon lifecycle extraction in #949, this closes the
remaining install-scanner findings the 0.5.x plugin was hitting. The
plugin source now contains neither process.env nor child_process; the
former moved to plugin config (resolved by OpenClaw before the plugin
loads), and the latter lives in @vectorize-io/hindsight-all under
node_modules where the scanner's directory walker skips it. The plugin
can be installed without --dangerously-force-unsafe-install.
BREAKING CHANGE: drops the llmApiKeyEnv plugin config field along with
the HINDSIGHT_API_LLM_*, HINDSIGHT_EMBED_API_*, and HINDSIGHT_BANK_ID
environment variables. Users must now configure llmProvider and
llmApiKey explicitly via 'openclaw config set'. Migration guide is in
hindsight-docs/docs-integrations/openclaw.md and the integration
changelog.
* chore(openclaw): pin published versions of hindsight-all and hindsight-client
Phase 2 (#949) introduced @vectorize-io/hindsight-all and
@vectorize-io/hindsight-client as plugin dependencies using 'file:'
workspace paths. Those paths resolve inside the monorepo but break when
the published tarball is installed outside it — 'openclaw plugins
install @vectorize-io/hindsight-openclaw' failed with 'Cannot find
module @vectorize-io/hindsight-all' because npm could not resolve the
file: path from the extracted extension directory.
Replace both with semver ranges targeting the published versions:
@vectorize-io/hindsight-all ^0.1.0
@vectorize-io/hindsight-client ^0.5.0
Verified end-to-end: 'openclaw plugins install <local-tarball>' now
succeeds without --dangerously-force-unsafe-install and without the
workspace-symlink hack. npm pulls both dependencies from the registry
into the extracted extension's node_modules, the plugin loads cleanly,
and 'openclaw plugins doctor' reports no issues.
Wires the Rust CLI up to every endpoint exposed by the Hindsight OpenAPI
spec and adds CI enforcement so new endpoints or new request-body fields
cannot slip in without matching CLI coverage.
Endpoints
- New `hindsight webhook {list,create,update,delete,deliveries}` and
`hindsight audit {list,stats}` subcommands.
- `hindsight bank` gains `set-disposition`, `consolidation-recover`,
`export-template`, `import-template`, `template-schema`.
- `hindsight memory` gains `history` and per-memory `clear-observations`.
- `hindsight document update`, `hindsight operation retry` added.
- Brings CLI coverage from 46/62 to 62/62 operations.
Request-body parameters
- Expose missing flags that the CLI was silently hardcoding: directive
`--priority`; mental-model `--tags` / `--max-tokens` /
`--trigger-refresh-after-consolidation`; recall `--query-timestamp`;
reflect `--fact-types` / `--exclude-mental-models` /
`--exclude-mental-model-ids`; retain `--document-tags`.
CI enforcement
- New `cli-coverage-check` entry point in `hindsight-dev` parses
openapi.json and verifies that (a) every operationId is called from
hindsight-cli/src/ (the progenitor client method names match the
operationId), and (b) every request-body property is present in
main.rs as a clap field or `long = "..."` attribute.
- Intentional non-exposures live in `hindsight-cli/.openapi-coverage.toml`
under `[skip]` / `[fields.<op>]` with a reason each (38 documented
field skips for flattened structs, nested structs, or fields surfaced
via a different subcommand).
- New `check-cli-coverage` job in .github/workflows/test.yml, triggered
on cli/core/dev/ci path changes, runs the script on every PR.
- smoke-test.sh exercises the new webhook / audit / bank-template /
set-disposition / consolidation-recover commands.
* feat: add HINDSIGHT_API_DEFAULT_BANK_TEMPLATE env var
Server-level default bank template applied automatically to every
newly-created bank. Holds an inline JSON BankTemplateManifest with the
same shape as the /import endpoint body. Fields set by the template
become per-bank overrides so they take precedence over equivalent
HINDSIGHT_API_* env defaults. The template is applied once on first
creation and never reapplied, so user overrides via PATCH /config are
never clobbered. Malformed manifests are logged and ignored so a broken
server-level setting cannot wedge bank creation.
* chore: regenerate docs skill
* test: update async_retain test mock for renamed bank_profile helper
* feat: add @vectorize-io/hindsight-embed daemon lifecycle package
Create a new top-level `hindsight-embed-npm/` package that owns the daemon
lifecycle for the Python `hindsight-embed` CLI: spawning via `uvx`, writing
the profile, waiting for `/health`, and shutting down. Nothing more.
Deliberately does not ship an HTTP client — `@vectorize-io/hindsight-client`
already covers retain / recall / reflect / createBank against the Hindsight
API, and the two packages compose: once `manager.start()` returns, consumers
talk to the daemon via `new HindsightClient({ baseUrl: manager.getBaseUrl() })`.
`HindsightEmbedManagerOptions.env` forwards an arbitrary `Record<string,
string>` to both the daemon process and the profile config via `--env K=V`,
and `extraProfileCreateArgs` / `extraDaemonStartArgs` escape hatches cover
any new CLI flag without waiting for a wrapper release.
Refactor `hindsight-integrations/openclaw` to consume both packages:
`HindsightEmbedManager` for daemon lifecycle in local mode, `HindsightClient`
for all HTTP memory operations. Drop the bespoke subprocess/HTTP client that
used to live in openclaw. The retain queue stays local to openclaw (it's a
client-side reliability workaround with a single consumer today — will move
to the client package or server-side when a second consumer needs it).
Wire the new package into the main release pipeline (versioned alongside
the other core packages, published from `v*` tags) and add a CI build job.
* docs: add Embedded Node.js SDK page for @vectorize-io/hindsight-embed
* refactor: rename hindsight-embed-npm to hindsight-all, restructure docs sidebar
The Node package previously named @vectorize-io/hindsight-embed was
semantically misnamed: hindsight-embed (Python) is a CLI tool, while what
this Node package actually provides is the Node equivalent of hindsight-all
— a programmatic lifecycle manager for a local Hindsight daemon. Rename to
match.
Package rename
- hindsight-embed-npm/ → hindsight-all-npm/ (git mv, history preserved)
- @vectorize-io/hindsight-embed → @vectorize-io/hindsight-all
- class HindsightEmbedManager → HindsightServer (matches Python hindsight-all)
- HindsightEmbedManagerOptions → HindsightServerOptions
- src/manager.ts → src/server.ts, src/manager.test.ts → src/server.test.ts
- openclaw (index.ts, backfill.ts, tests) and the claude-code Python port
updated to reference the new names
Docs restructure
- Split sdks/python.md: now client-only content. New sdks/hindsight-all.md
covers the programmatic hindsight-all Python package (HindsightServer and
HindsightEmbedded).
- Rename sdks/embed-npm.md → sdks/hindsight-all-npm.md with HindsightServer
examples.
- New "Installation" sidebar section, placed after Hosting, containing
Docker / Kubernetes / Bare Metal (anchor links into developer/installation)
plus Programmatic API (Python), Programmatic API (Node.js), and Daemon CLI.
- Add si-docker, si-kubernetes, si-nodedotjs, lu-hard-drive to the sidebar
ICON_MAP.
Docs dev-server fix
- docusaurus.config.ts: drop the flaky NODE_ENV sniff for including the
"Next" version. Use INCLUDE_CURRENT_VERSION exclusively. NODE_ENV was
unreliable across hot-reload paths and caused the Next version to
disappear intermittently when editing files.
- scripts/dev/start-docs.sh: export INCLUDE_CURRENT_VERSION=true so local
dev always shows Next; production builds leave it unset.
Lockfile cleanup
- package-lock.json and hindsight-integrations/openclaw/package-lock.json
had extraneous hindsight-embed-npm blocks left over from the rename.
Removed manually and verified with npm install.
* ci: fix openclaw jobs by pre-building workspace deps; regenerate docs-skill
The build-openclaw-integration and test-openclaw-integration jobs failed
with "Failed to resolve entry for package @vectorize-io/hindsight-all"
because openclaw depends on two monorepo workspaces via `file:` deps
(@vectorize-io/hindsight-client and @vectorize-io/hindsight-all) whose
`dist/` directories are gitignored and never built before openclaw's npm ci.
Both jobs now install the root workspace and build the two deps first,
mirroring the release-control-plane pattern.
Also regenerate skills/hindsight-docs/references/* via
./scripts/generate-docs-skill.sh:
- new skill pages for sdks/hindsight-all{.md,-npm.md}
- updated skill pages for sdks/embed.md and sdks/python.md to match
the new H1s and split content
- incidental refreshes to changelog/index.md, developer/models.md,
openapi.json, and uv.lock that verify-generated-files picked up
* ci: build openclaw before running tests so symlink test can realpath dist
PR #932 added update_mode (replace/append) to retain items but
did not update the docs. Add a section explaining the parameter,
when to use append mode, and a JSON example.
Closes#957
* feat(openclaw): add session pattern filtering for ignore and stateless sessions
Adds three new config options to the OpenClaw plugin that allow filtering
sessions by key pattern before recall and retain operations fire:
- `ignoreSessionPatterns`: glob patterns for sessions to skip entirely
(no recall, no retain). Useful for cron/scheduled agent sessions.
- `statelessSessionPatterns`: glob patterns for read-only sessions —
retain is always skipped; recall is also skipped when
`skipStatelessSessions` is true (default).
- `skipStatelessSessions`: boolean (default: true). When false, sessions
matching statelessSessionPatterns can still recall but never retain.
Pattern syntax mirrors lossless-claw: `*` matches non-colon characters,
`**` matches anything including colons. Session keys follow the OpenClaw
format `agent:<agentId>:<type>:<uuid>`.
Example config:
ignoreSessionPatterns: ["agent:*:cron:**"]
statelessSessionPatterns: ["agent:*:subagent:**", "agent:*💓**"]
skipStatelessSessions: true
Implementation:
- New `session-patterns.ts` module with compile/match utilities
- Session filter applied in `before_prompt_build` and `agent_end` hooks
immediately after the existing `excludeProviders` check
- New fields wired through `getPluginConfig`
- Schema added to `openclaw.plugin.json` (additionalProperties: false
was already set, causing config validation errors without this)
- 11 unit tests in `session-patterns.test.ts`
- 5 integration tests added to `hooks.integration.test.ts`
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* test(openclaw): support HINDSIGHT_API_TOKEN in integration tests
Pass HINDSIGHT_API_TOKEN env var through to HindsightClient and plugin
config in integration tests so tests work against authenticated APIs.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* docs(openclaw): document session pattern filtering options
Add ignoreSessionPatterns, statelessSessionPatterns, and skipStatelessSessions
to the README config table with glob syntax reference and usage examples.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Marco Rutsch <marco@rutimka.de>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* docs: add 0.5.0 release notes and changelog
* docs: include all commits since v0.4.22 and add recall perf to blog
* docs: include all commits since v0.4.22 and add recall perf to blog
* docs: add openrouter default model to provider table
* docs: reorder blog sections, fix code snippets, remove paperclip
* docs: add hermes integration docs link
* docs: fix broken anchor in blog post TOC
greenlet 3.4.0 lacks manylinux_2_41_aarch64 wheels. Use a UV_CONSTRAINT
file instead of the workspace lock file (which doesn't work in the
single-package Docker context).
Without the lock file, uv sync resolves fresh and picks up greenlet
3.4.0 which lacks arm64 wheels for manylinux_2_41, breaking the
multi-arch Docker build.
* fix: exclude local-llm from [all] extra to avoid heavy llama-cpp-python dep
local-llm (llama-cpp-python) requires C++ compilation and is only needed
for the built-in llamacpp provider. Keep it as a separate opt-in:
pip install 'hindsight-api-slim[local-llm]'
* feat: add local-llm optional extra to hindsight-all
Allows: pip install 'hindsight-all[local-llm]' to get built-in llamacpp support.
* chore: regenerate uv.lock from workspace root
* feat: add built-in llama.cpp LLM provider for fully local inference
Add `llamacpp` as a new LLM provider that manages a llama-cpp-python server
subprocess. Auto-downloads Gemma 4 E2B Q4_K_M (~3.5 GB) on first use and
runs inference locally via Metal/CUDA with no external services needed.
- New provider: `HINDSIGHT_API_LLM_PROVIDER=llamacpp`
- Singleton server shared across retain/reflect/consolidation
- Configurable: model path, GPU layers, context size, grammar enforcement
- User-extensible via `HINDSIGHT_API_LLAMACPP_EXTRA_ARGS`
- Flash attention + prompt caching enabled by default
- LLM provider cleanup on shutdown (stops subprocess)
- hindsight-embed: `--ui` flag on `daemon start`, removed FORCE_CPU on macOS
- Docs: configuration.md, models.mdx, providers grid updated
* chore: regenerate docs skill and update lockfile for local-llm dep
* feat: add update_mode='append' for retain to concatenate content to existing documents
When retaining with update_mode='append' and a document_id that already exists,
the new content is appended to the existing document text and the full document
is reprocessed. Delta retain automatically skips unchanged chunks, so only the
new content triggers LLM extraction.
- Add update_mode field to MemoryItem (API), RetainContentDict (internal), MCP tools
- Validate that update_mode='append' requires a document_id
- Fetch existing document content and prepend before processing in orchestrator
- Update Python, TypeScript, Go generated clients and top-level client wrappers
- Add tests for append, multiple appends, no-existing-doc, validation, and default replace
* fix: add update_mode field to Rust CLI and client MemoryItem initializers
* chore: regenerate docs skill references for update_mode
* docs: add best practice for filtering recall by memory shape (#856)
Add guidance on using entity labels with `tag: true` to deterministically
filter recall results when a bank contains different memory shapes
(e.g., concise rules vs. detailed procedures).
* feat: add OpenRouter support for LLM, embeddings, and reranking
OpenRouter is OpenAI-compatible for chat/embeddings and Cohere-compatible
for reranking, so no new provider classes are needed.
- LLM: added as OpenAICompatibleLLM provider (default model: qwen/qwen3.5-9b)
- Embeddings: reuses OpenAIEmbeddings with OpenRouter base URL (default: perplexity/pplx-embed-v1-0.6b)
- Reranker: reuses CohereCrossEncoder with OpenRouter rerank endpoint (default: cohere/rerank-v3.5)
- API key fallback chain: dedicated key → shared OPENROUTER_API_KEY → LLM_API_KEY
* chore: regenerate docs skill references and fix formatting
Extend format_facts_for_prompt() to include occurred_end and mentioned_at
temporal fields (when non-null), matching the MemoryFact model. Also add
RecallResponse.to_prompt_string() to Python and TypeScript client SDKs so
users can serialize recall results (with chunks and entity summaries) into
LLM-ready prompt strings.
Closes#924
* fix: make LiteLLM SDK embeddings encoding_format configurable (#925)
The hardcoded encoding_format='float' breaks providers like Voyage AI
(only accepts 'base64') and Gemini (doesn't support the parameter at all).
Add HINDSIGHT_API_EMBEDDINGS_LITELLM_SDK_ENCODING_FORMAT config option
that defaults to 'float' for backwards compatibility. Set to empty string
to omit the parameter for incompatible providers.
* chore: regenerate docs skill after configuration change
* security: bump lodash, lodash-es, and defu in root lockfile
Fixes Dependabot alerts in the root npm workspace lockfile:
- GHSA-r5fr-rjxr-66jc (high) lodash <4.18.1 (alert #338)
- GHSA-r5fr-rjxr-66jc (high) lodash-es <4.18.1 (alert #335)
- GHSA-737v-mqg7-c878 (high) defu <6.1.7 (alert #343)
defu (6.1.4 -> 6.1.7) and lodash (4.17.23 -> 4.18.1) were bumped via
targeted `npm update`. lodash-es was pinned exactly to 4.17.23 by
@chevrotain packages (transitive dep of mermaid in hindsight-docs),
so a `lodash-es` override (>=4.18.1) is added to the root package.json
to force resolution to the patched 4.18.1.
Verified: `npm ci` succeeds with 0 vulnerabilities. Mermaid/chevrotain
consumers all dedupe to lodash-es 4.18.1. lodash-es 4.x is semver-
compatible.
* chore: regenerate hindsight-docs skill
Picks up FAQ and best-practice sections added in #905 that were not
regenerated at merge time, so that `verify-generated-files` passes
for this branch.
* security: bump vite across integrations to patched versions
Fixes Dependabot alerts for vite transitive dev dependency:
- GHSA-v2wj-q39q-566r (high): server.fs.deny bypass with queries
- GHSA-p9ff-h696-f583 (high): related vite server vulnerability
Adds a `vite` entry to the npm `overrides` in each integration's
package.json to force the patched version (>=8.0.5). To make this
possible in ai-sdk, chat, and openclaw — which pinned vitest ^4.0.18
whose vite peer is `^6.0.0 || ^7.0.0` — the minor-compatible bump
vitest ^4.0.18 -> ^4.1.2 is also included. vitest 4.1.x supports
vite 8.x (peer: ^6 || ^7 || ^8), so all six integrations converge on
vite 8.x consistently.
paperclip had no overrides block; one was added.
Verified locally: `npm ci && npx vitest run` passes in all six
integrations (ai-sdk 23, chat 28, openclaw 66, opencode 89, paperclip 27,
nemoclaw 36 tests).
* chore: regenerate hindsight-docs skill
Picks up FAQ and best-practice sections added in #905 that were not
regenerated at merge time, so that `verify-generated-files` passes
for this branch.
Some LLM providers (e.g. Anthropic Haiku) return 1-indexed
content_index values. When only one content item is provided,
this causes KeyError: 1 since the dict only has key 0.
Clamp content_index to the valid range instead of crashing.
Fixes#873
Co-authored-by: easonysliu <easonysliu@tencent.com>
