Changelog

User-facing changes only. Internal maintenance and infrastructure updates are omitted.

This page covers the Hindsight core (API, CLI, control plane). Each integration is released on its own cadence and has its own changelog — find it on the integration's page.

0.9.2

Features

Recall callers can now supply an explicit temporal window instead of relying on parsed time expressions.·@nicoloboschi·3de41af86
Control plane now shows per-model staleness with consistent wording across views.·@nicoloboschi·97dbebb00
Automatic mental-model refreshes now enforce a minimum interval to prevent overly frequent refresh cycles.·@nicoloboschi·2b54a1d3c
Text-in embedding providers can now use asymmetric query/passage prefixes to improve retrieval quality.·@nicoloboschi·e256a7409
Knowledge-base tree now exposes each page’s refresh trigger, making refresh behavior easier to inspect.·@nicoloboschi·0dd32a3d6
Add a GitHub Copilot subscription provider, so Hindsight can run its LLM lanes on a Copilot plan.·@dudujuju828·1d9a6a381
Expose the knowledge base as native MCP tools, so an agent can browse, search and maintain its pages over MCP instead of falling back to HTTP.·@nicoloboschi·6582e26ef

Improvements

Retain processing uses less memory and handles large batches more efficiently by streaming sub-batch splitting.·@nicoloboschi·634312795
Mental model deltas are stored as verbatim Markdown blocks for more faithful, readable model change tracking.·@nicoloboschi·fe5c25d64
Batch retain is routed through batch-capable multi-LLM handling for better throughput and provider compatibility.·@handnewb·7f7d9c133
Database write overhead is reduced by disabling row counting when min_rows=0 and by interpreting 0 as 'off'.·@nicoloboschi·94815f5d0
Engine now serves addressed document/chunk/entity reads from the backing store, improving read consistency.·@nicoloboschi·7c5646436
Mental-model refresh operations now record what changes were made, improving auditability and debugging.·@nicoloboschi·300c546c5
CLI output is now non-interactive-friendly by disabling ANSI color and spinners when not running in a TTY.·@yossiovadia·b7c65fbb2
Vector index creation is now proportional to bank size instead of always creating three per bank, reducing DB bloat.·@nicoloboschi·edd0d0c5b
Retain and update_memory can now optionally resolve entities via a dedicated flag, enabling more controlled entity handling.·@nicoloboschi·df8ac42b5
Maintenance cross-tenant sweeps scale with tenant count to reduce unnecessary work on large deployments.·@nicoloboschi·3770d62e3
Worker polling uses a pooled DB connection per cycle with consolidated session setup, improving throughput.·@nicoloboschi·32b90cc98
The bank list endpoint is now paginated.·@nicoloboschi·6692e38c8
Background worker runs are now traced and join the caller's trace, giving end-to-end visibility into async work.·@nicoloboschi·08a119eeb

Bug Fixes

Consolidation now avoids retrying an entire batch when only some LLM calls fail, improving reliability of batch runs.·@nicoloboschi·ad6312a51
Retain now enforces a memory budget (rather than per-document limits) to prevent runaway memory usage on large inputs.·@nicoloboschi·e65973b1d
Consolidation claims are serialized per bank within a batch to prevent conflicts and inconsistent results.·@nicoloboschi·85bd23d50
Knowledge pages keep their intended scope reliably and trigger updates no longer reset it unexpectedly.·@nicoloboschi·99a319513
Entity resolution no longer merges entities based only on history-matched names, reducing incorrect merges.·@nicoloboschi·92695c6e5
Retain retries and fails clearly when fact extraction encounters schema drift, instead of silently producing bad data.·@nicoloboschi·d8a71c313
Reflect runs now fail if no answer is produced, preventing placeholder memories from being stored.·@nicoloboschi·f4a936950
Recall now skips over-budget facts instead of stopping early, and avoids returning empty matches.·@nicoloboschi·1f5006583
If llama.cpp is missing, the system now reports the configuration problem instead of looping on connection errors.·@nicoloboschi·bdee2bc88
Embedding profile env files are now read as UTF-8 to avoid failures with non-ASCII characters.·@koriyoshi2041·86675983b
Batch recovery is now correctly tied to the account that submitted the batch, preventing cross-account mixups.·@nicoloboschi·6ff6dc692
Provider content-policy refusals during retain are treated as permanent failures rather than endlessly retried.·@nicoloboschi·b81a3742b
Control plane: clearing the memories query no longer closes the search view unexpectedly.·@nicoloboschi·9ef1f1b84
Re-retain now correctly deletes chunks that were removed when content is reduced, preventing stale data from lingering.·@nicoloboschi·801f4910e
Query analyzer no longer misinterprets lone four-digit numbers as years.·@nicoloboschi·e20bb290f
Knowledge tree writes are serialized to prevent write conflicts and corruption under concurrency.·@Sanderhoff-alt·ea9efa860
CLI now exits with a non-zero status when configured with an invalid API URL, making failures script-friendly.·@Alan5168·71029fb81
Retain strict-schema behavior is now honored consistently for batch processing.·@kamilfurtak·2bba17ee4
Bank configuration is validated and persisted atomically to prevent partial or inconsistent config saves.·@nicoloboschi·2e0af7550
Consolidation now respects the configured temperature setting when generating consolidated results.·@lusoris·31cb1c2c8
Retain now fails closed on malformed memory-defense policies instead of skipping screening.·@nicoloboschi·400fd3684
Mental model refresh now removes references to facts that no longer exist, preventing stale citations.·@nicoloboschi·ba97f8585
Misconfigured retain settings now fail loudly instead of silently falling back to defaults.·@nicoloboschi·cc2dd3945
Ollama free-form calls now use the API path that honors num_ctx, improving context-window correctness.·@nicoloboschi·2409c2d5a
Voyage embeddings preserve asymmetric query/passage behavior to maintain retrieval quality.·@Tosko4·4d03f7d48
Knowledge-base page staleness is now computed per scope rather than from the bank watermark, improving accuracy.·@nicoloboschi·a5d00660b
Embedding configuration now inherits unspecified LLM settings instead of overwriting them with defaults.·@Sanderhoff-alt·e4e5f8b28
Recall budget limits now apply all the way through vector-index querying, improving predictability of results.·@nicoloboschi·31c1aaf21
Recall performance is improved by avoiding inefficient query plans on nested observation graph lookups.·@nicoloboschi·5f137bf39
Embedded daemon/UI lifecycle is hardened to reduce startup/shutdown errors and improve stability.·@nicoloboschi·d98990b46
Consolidation preserves source dates when observations are merged, improving timeline accuracy.·@nicoloboschi·2e2dfe130
Null metadata values are dropped during retain and recall to prevent invalid metadata from propagating.·@NovaLux12·19318ac08
Secret redaction uses ASCII token boundaries so secrets adjacent to CJK text are properly detected and removed.·@oldnicke·e181fb75c
Mental models now ensure only one refresh is queued at a time, preventing refresh backlogs and duplication.·@nicoloboschi·d69c53739
Knowledge-base page triggers are now patched (not replaced) on create/update, preserving existing trigger settings.·@nicoloboschi·b64943d19
List endpoints for mental models and directives now return a correct total count for pagination.·@nicoloboschi·803a45171
Automatic cache affinity is disabled for Azure OpenAI to avoid routing issues and instability.·@koriyoshi2041·7850efcef
Memory updated_at is now refreshed on all writes that change a memory, improving freshness tracking and sync behaviors.·@nicoloboschi·b919b97df
Retention sweep is now batched to prevent it from stalling foreground queries.·@nicoloboschi·0480f171e
Recall with long BM25 queries keeps the most selective terms, improving relevance and avoiding degraded search results.·@nicoloboschi·a37257ede
Retain writes keep entity postings and witness coverage consistent by applying them together in one write group.·@nicoloboschi·f660a4ad8
A mental model's last_refreshed_at records when the refresh finished again, not the source-data watermark, which is now carried by the new last_memory_seen_at. Clients testing staleness against last_memory_write_at should compare last_memory_seen_at.·@nicoloboschi·8fbdc6bf7
Maintenance skips a schema under concurrent schema change instead of deadlocking against it.·@nicoloboschi·435f1640b
Vertex AI's EU and US regional endpoints work again instead of returning 404.·@ferrastas·3e6a812f7
The embedded daemon no longer flashes console windows on Windows.·@Sanderhoff-alt·3f14672ea

0.9.1

Features

Add per-bank store capability checks, including the ability for a store to mark a bank as read-only (close it to writes).·@nicoloboschi·6c5857813
Allow selected HTTP request headers to be forwarded to extensions.·@rares-mollie·27d4386de
Add asynchronous document export for bank transfers.·@nicoloboschi·00b520e59
Improve bank import/export by preserving the Knowledge Pages tree and rebuilding related search state on import.·@nicoloboschi·fd294227c
Make all bank configuration fields exportable and importable via bank templates.·@nicoloboschi·3e3e3f372
Add a new xAI OAuth LLM provider for using LLM lanes with a SuperGrok subscription.·@JoshFunnell·82852c522
Add an opt-in forced-tool mode to produce structured outputs when using LiteLLM providers.·@nicoloboschi·7b35d2c6f

Improvements

Restore server-side prompt caching when using load-balanced OpenAI-compatible backends.·@JoshFunnell·e62015cbf
Speed up recall by making temporal extraction significantly faster without changing results.·@nicoloboschi·e8c42f2a7
Speed up recall scoring by evaluating observation expansion as sets instead of per-candidate rows.·@nicoloboschi·e6fb5d479
Improve recall efficiency by reusing entity IDs already returned in results to avoid redundant fetches.·@nicoloboschi·00acde015
Reduce large entity-pruning sweeps by processing entity pruning through a queue.·@nicoloboschi·681a79e6b
Make graph maintenance job submission smarter by deduplicating against running jobs as well as pending ones.·@cdbartholomew·a763109d8

Bug Fixes

Prevent recall from crashing when given extreme relative date offsets in queries.·@nicoloboschi·423e25db4
Ensure retrieved evidence is never dropped when synthesis is forced during reflection.·@nicoloboschi·07aec3b95
Fix document statements being incorrectly shared across banks by scoping them to the correct bank.·@nicoloboschi·da8077ade
Prevent lost turns when appending to the same document concurrently.·@nicoloboschi·11a4eda80
Fix batched file uploads so each file gets a unique storage key.·@nicoloboschi·45713adcc
Fix markdown table rendering in structured documents.·@dragonkid·e9a14da69
Fix knowledge base search errors on deployments without native text-search backends.·@nicoloboschi·18bff79ae
Fix Knowledge Pages search to use PGroonga when configured, improving correctness and reliability.·@JiehoonKwak·81b58934d
Add a DB-independent liveness probe so slow databases don’t cause pods to restart unnecessarily.·@nicoloboschi·00bad1711
Validate UUID inputs for entity and observation-history endpoints to prevent bad requests from triggering server errors.·@Alan5168·862a77c1a
Fix OpenAI/LLM reasoning handling by preserving literal inline <think> text and honoring configured reasoning effort.·@nicoloboschi·39ee97a61
Fix Llama.cpp requests so extra request fields (extra_body) are correctly passed through.·@handnewb·650da8c1e
Make Ollama “think” behavior configurable via extra request fields (extra_body).·@nicoloboschi·3a4b0214f
Prevent reranking failures/timeouts by batching FlashRank passage scoring instead of using an unbounded single pass.·@nicoloboschi·3e1b4611c
Prevent bank deletion deadlocks involving vector index schema changes during concurrent deletes.·@nicoloboschi·d733772e3
Validate bank configuration value types to prevent misconfigurations from wedging background tasks.·@nicoloboschi·2cf31d7fb
Improve transfer import handling by rejecting invalid import ZIPs with a clear 400 error explaining how to fix it.·@nicoloboschi·d506641dc
Fix whole-bank import to assign fresh internal IDs to avoid collisions.·@nicoloboschi·d7c33fdea
Fix document upload cleanup so failed uploads are correctly deleted.·@koriyoshi2041·78f1a0ef0
Avoid database lock issues by using a consistent lock order when inserting memory links.·@nicoloboschi·83a080f90
Fix long-running or stuck mental-model delta operations by tracing the call and decoupling its completion cap.·@nicoloboschi·500ab70f1
Fix consolidation after multi-round drains by refreshing all affected mental models.·@nicoloboschi·64ede3617
Fix consolidation batching so sibling tag groups are canceled when a batch fails, preventing cascading stuck work.·@nicoloboschi·dca002551
Fix pending consolidation statistics by removing memories that have permanently failed.·@nicoloboschi·dbe0ffb98
Fix embed daemon shutdown detection to use actual port occupancy instead of an unreliable health probe.·@Alan5168·c094ac27d

0.9.0

Features

Add per-bank toggles to enable/disable temporal search, graph expansion, and reranking during recall.·@cdbartholomew·d3946f17c
Allow configuring a per-budget cap on how many candidates the reranker considers during recall.·@nicoloboschi·359e619a4
Add a reranker failover chain so recall can fall back across multiple reranker configurations.·@nicoloboschi·76d8ba8a5
Add configurable locale detection for date parsing in the query analyzer.·@yufanw03·f2ae61eda
Add support for the OpenAI Responses API via a new openai-responses LLM provider.·@nicoloboschi·55883dc51
Add a generic per-input token cap for embeddings across providers.·@nicoloboschi·8e953c230
Allow per-operation overrides of extra LLM request body fields via configuration.·@nicoloboschi·cac55feda
Introduce Knowledge Pages: client-managed knowledge content with Control Plane UI and hindsight fs CLI management.·@nicoloboschi·218e6d34b
Add tools to troubleshoot mental model refreshes, including dry-run mode and trace retention.·@nicoloboschi·06e9c7054
Surface structured output for reflect/mental-model operations in the Control Plane.·@nicoloboschi·4278f0989
Introduce Hindsight Coding Agents: harness-pluggable long-term memory for coding agents.·@nicoloboschi·b5d8439c8
Add a local daemon mode for Coding Agents and let users choose the server during installation.·@nicoloboschi·e1b3d438e
Add --import-conversations to Coding Agents to migrate/import conversations from older per-agent plugins.·@nicoloboschi·17b7f46ae
Add a headless CLI tool to ingest/sync an Obsidian vault (hindsight-obsidian-sync).·@benfrank241·93fa0b016
Add a recall option to prefer returning observations (via OpenClaw integration).·@nicoloboschi·d19f54c77
Allow agent_knowledge_recall (Agent SDK) to request source chunks alongside results.·@chethanuk·cd4064939
Add support for custom OCR headers in the file parser.·@Sanderhoff-alt·f61d383ac

Improvements

Improve security and reliability of outbound webhooks by hardening delivery destinations and response handling.·@nicoloboschi·ad85affb1
Add Hermes-branded bank templates grounded in real Hermes user stories.·@benfrank241·5e557dbfc
Require an explicit Coding Agents install target (e.g., install all or a specific harness) to avoid accidental installs.·@nicoloboschi·05b55d20f
Make the memories storage backend pluggable so deployments can swap storage implementations.·@nicoloboschi·b769045b6
Add Control Plane indicators for documents being updated by an in-flight retain operation.·@nicoloboschi·3868c8d05
Show the originating coding agent’s logo on documents and memories in the Control Plane.·@nicoloboschi·aa38790dd
Improve the Control Plane UI design system and visual consistency.·@nicoloboschi·bf6c12d55

Bug Fixes

Fix bank restore/import so label entities are correctly classified.·@nicoloboschi·53948468a
Fix workers leaving behind in-progress operations by reconciling still-owned operations when a worker stops.·@nicoloboschi·c15b565c8
Prevent scheduled mental-model refresh jobs from enqueueing duplicate work.·@nicoloboschi·afdea53a9
Improve entity deduplication within a batch to avoid duplicate/variant entities being created.·@nicoloboschi·25a7237b1
Fix recalls so created_after/created_before time windows also constrain graph expansion.·@nicoloboschi·bda7ffd28
Fix embeddings so local providers honor configured query/document prompts.·@jpmf33·468cc4b7d
Fix LLM retry behavior to release concurrency permits during backoff to avoid unnecessary throttling.·@koriyoshi2041·f6b3ce3e3
Prevent unbounded log growth in the embedding daemon.·@koriyoshi2041·9d828cc7b
Keep observations in the original source language during consolidation.·@nicoloboschi·42e8c53da
Prevent mental model documents from being overwritten by delta-window candidates.·@nicoloboschi·ba5a4813b
Ensure internal (non-HTTP) recall paths also enforce the query token cap.·@nicoloboschi·5c15f28af
Return a proper 422 error when recall requests specify invalid fact types.·@tao943·4b76d8be2
Fix retention so metadata updates are applied even when memories are otherwise unchanged.·@Sanderhoff-alt·751deb47b
Fix auto-splitting of overlong retention output by unifying the OutputTooLong error handling.·@nicoloboschi·f572d8647
Show accurate last-write timestamps for banks and documents in the API.·@nicoloboschi·1ce308f35
Improve database connection handling so pooled connections aren’t held across embedding/LLM calls, reducing contention and timeouts.·@nicoloboschi·24825200b
Retry embedding connect timeouts for TEI-based embedding providers.·@jooneyp·417efb35d
Fix delete operations to sweep orphan entities even when no relink victims were enqueued.·@nicoloboschi·797faf798
Fix configuration errors to explicitly report which fields are disallowed when permissions are violated.·@kubaodias·13cacf21a
Fix Coding Agents session import attribution to use recorded working directory (cwd) rather than the session name.·@nicoloboschi·d238d2f7d
Fix Coding Agents Devin transcript ingestion by using node:sqlite and failing early if it’s missing.·@nicoloboschi·333812c85
Fix macOS installs so they work without requiring a Rust toolchain.·@nicoloboschi·8de576b4b

0.8.6

Features

Control plane: add document tag filtering and a unified facet chip UI for browsing memories.·@nicoloboschi·3a1841c42
Reflect: add an option to apply directives across all scopes (bypass directive tag scoping) when needed.·@nicoloboschi·6a460d2c9
Retain: detect and report documents that produce zero extracted facts at write time.·@nicoloboschi·9452ac29d
Configuration: allow per-operation overrides of LLM reasoning_effort.·@nicoloboschi·b1a0ef5f7
Temporal parsing: add Russian temporal period rules for better date/period understanding.·@dimonnld·4724f26d3
Configuration: make embedding thresholds configurable.·@Sanderhoff-alt·ada3329bb
Control plane: display the API version in the sidebar.·@handnewb·7a9ea7058
Engine: add a bulk API to delete multiple memory units in one operation.·@chrislatimer·64fe5e81f
API/UI: filter memory lists by linked entity and add an entity timeline view.·@nicoloboschi·c41ad9bd7
Extensions: support extension-owned, bank-scoped tables with automatic provisioning.·@nicoloboschi·57c18bc29
Configuration: allow per-bank overrides for storing document text (store_document_text).·@nicoloboschi·6a0b85f10
TypeScript client: support custom HTTP headers on requests.·@menelsystemsbot·dbf3b9d9b
Engine: add created_before filtering to list_memory_units to query memories by ingest age.·@nicoloboschi·441cf2272
Observability: a stalled /health check now self-diagnoses — it reports whether the event loop is blocked or the database connection pool is exhausted.·@nicoloboschi·489d55fa6
Added a GitHub Copilot CLI integration that recalls memory before a session and retains a summary after.·@ScottGuymer·6500944c7

Improvements

Reflect: fail fast on unusable tool calls instead of attempting to salvage leaked text, improving correctness and safety.·@nicoloboschi·678ca0e90
Performance: reduce redundant graph seeding and UUID scans for faster operation on large datasets.·@zommiommy·b475f5cca
Retain: reject misplaced file retain metadata to prevent invalid ingestion requests.·@r266-tech·1fa2de332
Workers: per-type slot settings are now named HINDSIGHT_API_WORKER_<TYPE>_RESERVED_SLOTS, since they reserve a guaranteed minimum rather than cap concurrency. The old *_MAX_SLOTS names keep working but log a deprecation warning.·@nicoloboschi·979999651
Python 3.14 is now supported.·@nicoloboschi·a514d3962

Bug Fixes

Control plane: dark mode styling now correctly follows the configured dark theme class.·@nicoloboschi·94f4adfbe
Python and other clients: expose mental-model query controls (and async retain operation_id) so callers can tune queries and track async operations.·@nicoloboschi·70c09adcf
Reflect: mental-model scoping is now applied correctly for empty scopes.·@Sanderhoff-alt·c5643fdf5
Graph maintenance: fix a race between enqueuing maintenance work and worker draining to prevent missed/duplicated queue work.·@nicoloboschi·40d2b7f6b
Oracle support: fix full-text/vector and audit/log related SQL issues to improve compatibility and reliability on Oracle.·@dcbouius·4f4c2988e
Tracing/observability: provider responses are now safely serialized even when unvalidated.·@oldnicke·b8e1524a1
TEI embeddings/reranking: automatically retry on HTTP 429 backpressure to reduce transient failures.·@r266-tech·91160f3ba
Workers: prevent wedged retain jobs from occupying worker slots indefinitely.·@nicoloboschi·8133c5ab7
LLM integrations: reasoning_effort is now consistently sent across tool paths/providers (including Codex).·@merlinr68·9a1ba951f
Batch retain: improve retry/recovery so parent jobs can be re-run and stranded work is reconciled after failures.·@nicoloboschi·ac4df7eb8
Transfers/import: preserve consolidation lifecycle state when importing an entire bank.·@nicoloboschi·af196287e
Local embeddings/reranking: default to CPU on Apple Silicon to avoid an MPS memory leak.·@nicoloboschi·2620a2a3f
Claude Code integration: reflect tool calls now work and honor the configured model.·@nicoloboschi·8f19087c2
Control plane: audit-log and observations tabs now respect resolved per-bank configuration, and observations inheritance is preserved.·@nicoloboschi·ed248447e
Library import: avoid dotenv side effects when importing Hindsight as a dependency.·@nicoloboschi·5792b2b86
Consolidation: stop emitting an unsupported maxItems field that broke AWS Bedrock consolidation.·@r266-tech·c908fade1
Retain/dry run: return free-form entities from dry-run extraction and accept a text alias in fact extraction inputs.·@oldnicke·f7ff5341f
Graph: edited/restored memories are now re-queued for relinking so links stay up to date.·@Sanderhoff-alt·73b575c7a
Curation: preserve causal links across edit and invalidate/restore operations.·@nicoloboschi·0e5aa8896
Auth: ensure bank writes are authorized before provisioning occurs.·@Sanderhoff-alt·0f47c7a8d
Helm/Kubernetes: ensure workers use the correct API port override to avoid service-discovery failures.·@Fyko·d29f4703e
Mental models: fix delta refresh watermarking so refreshes anchor to the newest processed memory.·@nicoloboschi·af8cf142d
Retain: async retries are now idempotent via caller-supplied operation_id.·@nicoloboschi·31218127e
CLI: preserve HTTP 400 error details to make failures easier to diagnose.·@koriyoshi2041·1ff09ccf9
LLM endpoints: normalize LM Studio/Ollama base URLs to /v1 for compatibility.·@nicoloboschi·47a7d4380
Retain/linking: fix within-batch cosine similarity and adjust causal-target offsets to improve link quality.·@koriyoshi2041·7b161740d
Backups: fix restore schema compatibility issues.·@oldnicke·3cc371382
Retain: oversized append-only documents are recognized as appends again, so re-ingesting long histories no longer re-extracts everything from scratch.·@Korayem·a6c875156
Retain: fix a database deadlock where flushing statistics took a second connection while the first still held uncommitted row locks, which could hang a retain indefinitely (notably on Oracle, where no deadlock error is raised).·@nicoloboschi·6d5157575

Other

Zapier: remove the memoryDefenseTriggered trigger (gated capability) from the integration surface.·@benfrank241·e57765e01

0.8.5

Features

Add opt-in request dumps for rejected (4xx) LLM calls to make diagnosing provider errors easier.·@nicoloboschi·be6caf9dc
Add an optional fail-on-extraction-errors mode for retain(), allowing stricter handling when fact extraction fails.·@benfrank241·ec3b415c4
Add tag-based filtering to memory listing endpoints (list_memories and list_memory_units).·@superafun·c1fadc008
Allow deleting terminal bank operations via the API (with control-plane support).·@chethanuk·dcef72480
Allow overriding audit logging settings per bank.·@nicoloboschi·0d2dbe756
Add support for pg0:// URLs carrying optional user/password credentials for easier configuration.·@nicoloboschi·0c38d46ee
Add a configurable cap on planner parallelism to reduce connection pool pressure on busy deployments.·@cdbartholomew·347b9c23c
Support Anthropic Message Batches (batch API) for discounted token usage.·@hb-cam·7f2df54e0
Add Anthropic prompt caching support (including carrying cache markers in batch system prompts) to reduce repeated prompt costs.·@hb-cam·7c3a5619f
Add step-by-step context caching for Gemini tool loops to improve efficiency across multi-step tool interactions.·@nicoloboschi·81aa4979b

Improvements

Add bank attribution for remote reranker calls to improve per-bank observability and accounting.·@Jordan-Jarvis·8a7a70b82

Bug Fixes

Recover from malformed LLM JSON responses using a last-resort repair fallback, reducing failures on structured outputs.·@nicoloboschi·a23187a45
Ensure per-operation strict JSON schema handling works correctly and respects explicit per-call opt-out.·@nicoloboschi·9fe339dfb
Improve compatibility with OpenAI-style tool calling by emitting canonical tool-result messages and preserving required tools for custom endpoints.·@Jordan-Jarvis·434dbee64
Fix Claude Code LLM calls with max_turns=1 by preventing built-in tools from interfering with single-turn requests.·@peter216·b82fb603c
Increase reliability of retain() fact extraction by handling retries/edge cases (zero retry budgets, boundary batches, degenerate text) and preserving fact alignment.·@chethanuk·912f8e22d
Make retain() more robust by reasserting resolved entities before linking and standardizing extracted entities as plain strings.·@nicoloboschi·0679d38e8
Preserve bank attribution and refresh search vectors during memory curation, improving correctness of curated memories.·@Jordan-Jarvis·6ba98c040
Fix audit coverage so memory updates are properly recorded when auditing is enabled.·@handnewb·000fb9ddb
Repair vector index coverage per bank after restore/upgrade to prevent missing or incomplete search results.·@nicoloboschi·bd853be35
Improve stability and correctness of background operation processing (retry budgeting, completing successful operations, bounding terminal history, and isolating side-effect failures).·@handnewb·234f5a062
Fix delta refresh watermark handling so no-op refreshes still advance and don’t cause repeated processing.·@SunneeYang·c1fae2ae1
Fix PostgreSQL startup to honor credentials embedded in the database URL.·@r266-tech·820019f3c
Fix database migrations to correctly handle percent-encoded database URLs and install maintenance routines in the configured schema.·@Jordan-Jarvis·8bd9ce194
Improve date parsing by selecting the strongest match and fixing day+month+year interpretation (avoids collapsing to whole-month dates).·@benfrank241·d28b85273
Fix crashes and incorrect results in Chinese temporal extraction (including year-0 edge cases).·@r266-tech·5f1a86765
Prevent oversized embedding requests from failing by truncating inputs before embedding.·@nicoloboschi·263eba134
Improve consolidation robustness by requesting JSON dedup decisions and normalizing/handling invalid dedup actions safely.·@r266-tech·6d82c8a55
Fix memory browsing endpoints to return associated metadata.·@Sanderhoff-alt·5cc1482a7
Preserve JSONB and timestamp provenance during transfers to avoid losing data fidelity.·@Korayem·d7c32f963
Make LLM trace writing safer by avoiding writes during startup/shutdown races and gating on lifecycle state.·@nicoloboschi·3fb33b987

0.8.4

Features

Added Requesty as an OpenAI-compatible LLM provider option.·@Thibaultjaigu·6eb85570a
Recall can now return structured per-stage scores and supports two-level min_scores filtering for finer result control.·@nicoloboschi·758f346d3
Mental models can now be refreshed on a cron schedule automatically.·@nicoloboschi·dae18b1fa
Multi-LLM setups now support configuring LiteLLM Router and Vertex service account keys.·@nicoloboschi·47992d843
Multi-LLM configurations can now set Vertex AI project and region per indexed member.·@cdbartholomew·4fe477eaa
Added multi-LLM failover and round-robin selection via indexed configuration.·@nicoloboschi·f7c7a62e5
Added Atlas Cloud as an OpenAI-compatible LLM provider option.·@lucaszhu-hue·f8d277697
Recall can now prefer observations to avoid returning raw facts that have been superseded.·@nicoloboschi·5f0b71551
Recall supports configurable recency decay (linear, exponential, or none) to tune ranking toward recent memory.·@nicoloboschi·a2166ee4f
Local reranking can now accelerate on Intel XPU where available.·@r266-tech·1a09a9ccc
Added an Aider integration to bracket sessions and persist memory more cleanly.·@DK09876·aab703207
Added an Eve (Microsoft Agent Framework) MCP connection helper for easier integration.·@benfrank241·d0b77f5be
Added a Windsurf (Codeium) integration via MCP.·@DK09876·7194f98b1
Added a GitHub Copilot (VS Code) integration via MCP.·@DK09876·91bf32842

Improvements

Memory explorer now shows full memory details for easier inspection and debugging.·@koriyoshi2041·f8ce15b9b
Bank stats are faster and more scalable, with a refresh option to force recomputation when needed.·@nicoloboschi·d68f61896
LLM configuration now supports per-operation temperature control for more predictable behavior across different workflows.·@nicoloboschi·a27754fb1
Reflect structured output will retry fewer times, reducing latency and avoiding runaway retries.·@koriyoshi2041·85599f3ef
Standalone workers now warn when worker_id is missing or unstable, helping prevent misconfigured deployments.·@nicoloboschi·680305aea
Recall tracing now reports phase timing that matches total request duration more accurately.·@nicoloboschi·9e06237e4
Token usage reporting now includes cached tokens and model “thoughts” tokens where available.·@cdbartholomew·b4874672f
MCP responses omit reflect directives_applied by default to reduce noise in traces.·@r266-tech·246803bcf
Control plane upload size limit is now configurable.·@nicoloboschi·5543992d7
Database migrations are faster and more flexible, including optional extension reconciliation and dropping an unused global vector index.·@nicoloboschi·51cb32896

Bug Fixes

Batch fact extraction now honors the configured LLM temperature, making outputs consistent with your settings.·@r266-tech·8ed49e438
CLI explore now correctly handles header selection when displaying results.·@koriyoshi2041·82afa7618
Per-scope LLM timeouts and retry policies are now correctly applied to providers.·@nicoloboschi·40fe7aac8
LiteLLM-backed providers now correctly apply configured default request headers.·@r266-tech·5e73d5ff6
Dynamic banks now correctly inherit configured defaults.·@devinkuhn·ef2e8ab7f
Reflect now correctly parses JSON-wrapped responses from models.·@xmh1011·82b01ace5
Trace-based token/cost accounting is preserved even when tool-call argument parsing fails.·@nicoloboschi·01eda5188
Python and TypeScript SDK wrappers now correctly pass recall min_scores filtering options.·@r266-tech·cce0a2cb3
Consolidation no longer drops items when a dedup action is missing; it defaults to keeping data.·@koriyoshi2041·1c1cf4ce5
Markitdown parser now correctly handles certain UTF-8 text files that begin with an ASCII prefix.·@nicoloboschi·74e82a3ea
Database migration bootstrap now respects the configured vector extension.·@r266-tech·017b8d727
Consolidation now keeps search vectors up-to-date for observations, improving recall quality and reliability.·@qxxaa·21176f8ee
Embedding on Windows now reliably finds pythonw when installed alongside the API script.·@r266-tech·911d27fc5
Retain append mode now correctly merges JSON arrays to preserve conversation-aware chunking.·@qxxaa·78d32cd16
Control plane now shows documents that are uploaded but still pending processing.·@nicoloboschi·91e095afa
PATCH bank no longer accidentally creates new banks.·@Sanderhoff-alt·a7d1f26f9
Dry-run extraction no longer creates banks as a side effect.·@Sanderhoff-alt·0673131a8
Memory defense now shows the correct number of matched patterns.·@Sanderhoff-alt·8e66c397a
CLI pagination now correctly supports large limit/offset values.·@nicoloboschi·63a92bef5
Gemini structured output prompts no longer duplicate schema instructions.·@r266-tech·199ae146a
List endpoints now reject negative limit/offset with a clear 422 error instead of a server error.·@r266-tech·4db8a1236
Reduced risk of graph maintenance queue deadlocks under concurrency.·@cdbartholomew·cabcb3bb0
Recall can now accurately filter for untagged/global observations.·@nicoloboschi·0ba613c3c
Async operations now return a clean 404 for missing banks instead of an internal database error.·@cdbartholomew·04703d215
Bank configuration writes now validate disposition_* ranges to prevent invalid settings.·@r266-tech·387c09e91
Anthropic strict structured output is now produced more reliably by using tool_use instead of prompt-based injection.·@eldar702·cbce93704
Bank stats no longer become stale after deletes/clears thanks to improved cache invalidation.·@r266-tech·f21944d78
Bank stats cache invalidation no longer leaves stale results in some cases.·@zupengwang·ccd3eb24c
Retain chunking is now idempotent, preventing failures when structured chunk size settings increase.·@nicoloboschi·ba5ddd59a

0.8.3

Features

Added OpenHands integration with native MCP configuration and built-in recall/retain rules.·@DK09876·65862c4fe
Added Zed editor integration via an MCP context server with recall/retain rules.·@DK09876·539101af3
Added Continue.dev integration via an HTTP context provider.·@DK09876·c05ab9103
Added Composio integration so Hindsight memory can be used as custom tools.·@benfrank241·1c9ba0e65
Added a Zapier app with actions and REST-hook triggers for automating workflows with Hindsight.·@benfrank241·359b2bc76
Added a dry-run fact extraction API endpoint to preview extraction results without saving them.·@nicoloboschi·d382b340f
Added optional MarkItDown-based OCR support for extracting text from documents.·@Sanderhoff-alt·5ee53c512
Improved Chinese temporal query understanding for more accurate time-based searches.·@Sanderhoff-alt·d0c54560a
Added tool annotations (read-only/destructive hints) to MCP tools for safer client behavior.·@yunanwg·27aa6bbf4
Added support for the "shared" observation scope across consolidation and the Control Plane UI.·@nicoloboschi·0a046f97a
Added support for exact tag matching (tags_match=exact) across the UI and client libraries.·@nicoloboschi·2f075dede
Added Intel XPU detection to enable local embedding acceleration when available.·@nicoloboschi·2fb2de1aa

Improvements

Exposed additional operation/worker backlog metrics to make async processing health easier to monitor.·@MattJackson·b1fe23fbe
Added Gemini service tier configuration options.·@r266-tech·d68bd0742
API responses now omit wire-safe null fields to reduce payload size.·@nicoloboschi·e20a36c95
Improved search recency scoring using effective-time fallback for better ranking of results.·@r266-tech·3d6b19af5
Parallelized tenant schema migrations to reduce migration time on multi-tenant deployments.·@nicoloboschi·460fc63d9

Bug Fixes

Fixed retain chunking so oversized documents no longer lose content when split into multiple sub-batches.·@nicoloboschi·81865bf87
Fixed MCP so custom tool instructions are applied correctly again.·@Sanderhoff-alt·f4bac2d41
MCP responses no longer include reflect tool/LLM traces by default, preventing unexpected extra output.·@yunanwg·44972d321
Prevented hung LiteLLM calls from blocking forever by enforcing a hard timeout.·@kubaodias·551932991
Fixed bank listing so bank-config disposition and mission overlays are applied consistently.·@PathKnower·cb73790c2
Dry-run extraction now rejects empty/whitespace inputs and runs only after passing the precheck step.·@r266-tech·a92e25bcf
Fixed consolidation when the LLM returns single-value source_fact_ids so runs don’t fail unexpectedly.·@haodonp·94619ce52
Fixed metric labeling for HTTP endpoints so bank IDs are templated correctly (avoids high-cardinality labels).·@sireika·4efa20472
Fixed provider quota reset behavior by deferring retries to avoid unnecessary retry storms.·@oldschoola·2b521c3a0
Honored CODEX_HOME for locating Codex auth.json, fixing auth in custom home directories.·@mdenil·0c9bc765c
Fixed the Control Plane so the “shared” observation scope is selectable in the Add Document UI.·@r266-tech·0ca0226d3
Improved output sanitization by stripping malformed reasoning tags/blocks from non-structured outputs.·@grimmjoww·7b17da7a0

0.8.2

Features

Added support for observation scopes, including listing, filtering, visualization, and per-scope observation limits.·@nicoloboschi·b08f43496
Added SIEM-focused enrichment fields to Memory Defense webhook events to improve downstream security alert context.·@chrislatimer·f304ce7e0
Introduced reversible memory curation so memory units can be edited, invalidated, and reverted.·@nicoloboschi·de22b606e
Added per-bank provider cost attribution using the OpenAI "user" field.·@sireika·fb16fc4fd
Added a native Nous Portal LLM provider with OAuth authentication (no external CLI dependency).·@nicoloboschi·82c7df726
Added a per-bank LLM connectivity probe to help verify provider configuration.·@nicoloboschi·0d55a9b78
Added Gemini Batch API support for retain fact extraction.·@nicoloboschi·a942b1c81
Added support for the Gemini embedding-2 model family (per-input embeddings).·@nicoloboschi·22ae72a90
Improved retain ingestion by chunking JSONL files on line boundaries to avoid splitting records.·@nicoloboschi·4e7780e59
Added an option in the UI to export a constellation as a shareable SVG poster.·@nicoloboschi·0280b3486
Added a Cursor plugin providing Hindsight persistent memory integration.·@DK09876·91d767cdc
Added a pip-installable Cline integration package (hindsight-cline).·@benfrank241·36e31c675
Added a pip-installable Cursor CLI integration package (hindsight-cursor-cli).·@benfrank241·c6dd08944
Added Hindsight memory integration for Microsoft Agent Framework (no MCP required).·@benfrank241·f5a6c300f

Improvements

Added a local Control Center web app for embedded deployments.·@nicoloboschi·4032b2791
Improved reflect synthesis to better respect directives and language rules in final outputs.·@nicoloboschi·90ee101be
Added support for selecting Amazon Bedrock service tier.·@mobilinkd·27cb1c684

Bug Fixes

Fixed the control-plane bank view to avoid double-loading graph data, improving performance and reducing redundant requests.·@nicoloboschi·d81486ff9
Fixed per-scope observation limiting so a limit of 0 correctly prevents new observations from being created.·@r266-tech·f6710963e
Improved cancellation handling so abandoned recall/reflect operations can be reliably cancelled on disconnect.·@nicoloboschi·07c85da98
Fixed a concurrency issue where model initialization could change the global Torch default dtype, ensuring consistent float32 behavior.·@nicoloboschi·30fb287d1
Fixed webhook payloads to include previously missing fields.·@chrislatimer·87448b161
Fixed PostgreSQL history storage by widening history bank_id to TEXT to prevent failures with longer IDs.·@nicoloboschi·621ab7e66
Fixed the CLI to correctly parse get-memory responses.·@nicoloboschi·72d9881a6
Fixed the control-plane meeting notes flow to drop empty messages that could fail locale validation.·@nicoloboschi·8e6dc5fcd
Improved rolling-deployment migration skipping to handle wrapped dependency-resolution errors more reliably.·@bionicbutterfly13·0afa046fc

0.8.1

Features

Add an API configuration flag to skip storing raw document text.·@nicoloboschi·1296e9fc1
Add an Oh-My-OpenAgent (OMO) integration for connecting Hindsight memory to OMO workflows.·@dcbouius·6dc56498c
Add a Cline integration that hooks into lifecycle events to save and retrieve Hindsight memory.·@benfrank241·66e58a23a
Add a Haystack integration to use Hindsight as a memory component in Haystack pipelines.·@DK09876·394d66e60
Add a Cursor CLI integration for capturing and using Hindsight memory from Cursor workflows.·@Korayem·dbfe83a2a

Improvements

Introduce an official Obsidian plugin to sync and use Hindsight memory from Obsidian.·@benfrank241·b0f86f9c0
Package the Roo Code integration as an installable PyPI CLI (hindsight-roo-code).·@benfrank241·644e37ac1

Bug Fixes

Fix vector index behavior to avoid forcing vchordrq.probes settings on vchord indexes that don't support it.·@nicoloboschi·109e1bd95
Ensure database maintenance routines are installed correctly when using the public schema.·@nicoloboschi·95d77233b

0.8.0

Features

Add periodic background maintenance to reconcile consolidation state and enforce retention across tenants.·@nicoloboschi·e77461762
Add durable progress snapshots for long-running consolidation and batch retain operations, so status can be resumed/inspected reliably.·@nicoloboschi·c94935bfa
Enable LLM request tracing by default (with per-bank support) and retain traces for 1 day for easier debugging/observability.·@nicoloboschi·f4a3329ce
Add semantic deduplication of near-duplicate observations during consolidation (enabled by default, except on Oracle).·@nicoloboschi·8aa31edd4
Store mental-model and observation history in dedicated database tables for better scalability and reliability.·@nicoloboschi·7e1145c08
Add a local ONNX embeddings provider so embeddings can be generated without a hosted embeddings API.·@shoveller·b5a324b77
Add whole-bank export/import for cross-instance migration.·@nicoloboschi·602c9f55e
Add document export/import between banks without re-running the LLM (reusing previously extracted results).·@nicoloboschi·1d6d73bce
Add provider-agnostic prompt-prefix caching (default on) to reduce repeated LLM work for retain, consolidation, and reflect.·@cdbartholomew·d7a3aa526
Add global support for sending extra provider request fields via HINDSIGHT_API_LLM_EXTRA_BODY.·@nicoloboschi·01296d8d5
Honor HINDSIGHT_API_LLM_STRICT_SCHEMA across all JSON-schema capable LLM providers for stricter structured outputs.·@nicoloboschi·a3d3d42b3
Add Superagent safety middleware integration for safer agent execution.·@DK09876·b70830218

Improvements

Improve client SDK trace visibility by exposing reflect tool/LLM call details in the Python and TypeScript wrappers.·@nicoloboschi·0db70bb88
Allow tuning recall behavior with configurable per-strategy retrieval boosts via environment configuration.·@nicoloboschi·01134047d
Allow tuning semantic recall sensitivity with a configurable minimum similarity threshold (HINDSIGHT_API_SEMANTIC_MIN_SIMILARITY).·@zwcf5200·aa024a5cd
Avoid an unnecessary extra LLM call during reflect when a fresh mental model can be used directly.·@nicoloboschi·c255d3552
Speed up and reduce load from bank statistics queries via caching and more efficient computation.·@cdbartholomew·7683f2900
Improve visibility when audit logs or LLM requests are disabled by showing a clear "not enabled" splash in the control plane UI.·@nicoloboschi·283419280
Improve OpenCode integration observability with clearer debug logging, endpoint reporting, and surfaced errors.·@nicoloboschi·796a9eff9

Bug Fixes

Fix consolidation runs failing or truncating output by enforcing an explicit output token budget.·@xmh1011·c2524473e
Reduce slow/expensive temporal recall paths by limiting how many entries are scanned per fact type.·@nicoloboschi·4c33a4e55
Fix pgroonga BM25 search failures by escaping query text properly.·@xmh1011·6b8fc53d7
Preserve correct RRF (fusion) source ranks in retrieval traces/results.·@zwcf5200·f62500193
Prevent the API from hanging indefinitely at startup by failing fast when model initialization blocks.·@nicoloboschi·4f5003480
Fix recall and mental-model history support when using the Oracle database backend.·@DK09876·23710f4a8
Validate embedding vector dimensions before writing to pgvector to prevent storage/query errors.·@ai-ag2026·06c88e043
Improve retain concurrency safety by rechecking freshness before extraction and serializing concurrent writes to the same document.·@nicoloboschi·23168ebf6
Fix incorrect fact attribution during retain caused by bank routing key leakage.·@nicoloboschi·d695611ad
Fix mental model creation failures by ensuring the bank exists before inserting the model.·@r266-tech·56b4271d9
Fix external Postgres setups using vchord by adding required catalogs to the session search_path.·@nicoloboschi·049901802
Improve Docker standalone diagnostics by providing a clear error when pg0 bind-mount permissions prevent startup.·@nicoloboschi·75a7c19d6
Improve compatibility with LLM servers that mishandle tool_choice="required" by falling back to a safer tool-choice mode.·@nicoloboschi·e30f8af14
Fix Bedrock IAM authentication with the LiteLLM reranker by making the reranker API key optional.·@r266-tech·3c8ca47dd
Expose retain outcome metadata so users can see what happened during a retain operation.·@xmh1011·831f0efa1
Make bank config PATCH persist correctly even for banks that have never retained data yet.·@nicoloboschi·a809547aa
Prevent fact extraction prompt formatting from breaking extraction by removing Markdown bold markers.·@Oxygen56·50b7eda2a
Improve recall stability/performance by gating VectorChord BM25 usage and applying per-source candidate caps.·@nicoloboschi·70d98c7a2
Fix control plane UI localization issues for operations and graph legends.·@MapleEve·a14ce623c
Improve transfer import reporting by including mental_model_history counts in CLI summaries.·@r266-tech·3346363d2

0.7.2

Features

Added Google ADK integration to connect Hindsight memory with Google ADK agents.·@benfrank241·c032a74f
Added Flowise integration with Hindsight memory tools.·@benfrank241·41ad2b55
Added Fireworks AI as a batch inference provider.·@Fyko·537b2812
Added Roo Code integration via MCP (including rules support) to connect Roo Code with Hindsight memory.·@benfrank241·bf6b9026
Added Vapi webhook integration to send voice-agent interactions into Hindsight memory.·@benfrank241·dde133da
Added Gemini Spark integration via MCP.·@benfrank241·7cdacc4b

Improvements

Improved graph maintenance performance for vector search seeding (faster ANN maintenance work).·@nicoloboschi·99c7367f
Improved vector search correctness/tuning by using cosine opclass for vchord ANN and applying backend-specific tuning settings.·@isac322·e4686b92
Reduced consolidation retry backoff to recover faster from transient failures.·@nicoloboschi·f49b85c0

Bug Fixes

Fixed a control-plane i18n redirect loop by pinning the Next.js version.·@nicoloboschi·201f5d7c
Limited native ML thread pools to available CPU cores to prevent resource overuse and instability.·@nicoloboschi·0a2ee845
Fixed PostgreSQL upgrade issues for v0.7.x deployments.·@nicoloboschi·b7f267b0
Fixed backup/restore to include previously missing tables.·@nicoloboschi·867b7b4a
Made memory consolidation more reliable under concurrency by improving scope-aware locking/dispatch and atomic submission behavior.·@nicoloboschi·08ce8176
Fixed retention of oversized documents to prevent incorrect chunk indexing and loss of full document content.·@nicoloboschi·364ccf17
Improved retain/recall robustness for unusual Unicode inputs (special-token literals and lone surrogates).·@nicoloboschi·4bc7013e
Fixed a race condition when inserting memory links during retain operations.·@nicoloboschi·8cf0dcbf
Fixed Windows embedded daemon launch behavior to avoid opening an extra terminal tab.·@nicoloboschi·4280ac3f
Improved embedded Postgres stop/restart reliability by updating the embedded DB dependency.·@nicoloboschi·8d9000a8
Fixed CLI memory retain timestamp handling and corrected supported fact-type values.·@slayoffer·df73c792
Fixed directive listing/reflect to honor tag groups so filtering behaves as expected.·@nicoloboschi·fb554664
Fixed daemon networking to honor an explicitly configured host and port.·@Sanderhoff-alt·77380211
Isolated the Claude Code provider subprocess from user plugins to prevent interference and improve reliability.·@nicoloboschi·bcae23d9
Improved daemon startup reliability by waiting for health before reclaiming its port.·@s09x·5e547f71
Made worker retry behavior respect the configured maximum retries.·@aaronwestphal·85f6769e
Prevented retain from silently dropping memories when fact extraction fails.·@nicoloboschi·6e734e1a
Fixed control-plane auth redirects to validate return targets and honor basePath (improving security and correct routing).·@nicoloboschi·9571a341
Preserved raw reranker scores for providers that already return calibrated [0,1] scores.·@nicoloboschi·18b9c596

0.7.1

Features

Allow configuring LLM "reasoning effort" via environment variable.·@s9rkn·1890d2b7
Add per-provider HTTP timeout environment variables for reranker requests.·@nicoloboschi·a510b07a
Add additional Chinese locale variants in the control plane UI.·@MapleEve·617939d8

Improvements

Schedule consolidation work using bank priorities so higher-priority banks are processed first.·@nicoloboschi·cf637799
Expose "graph_maintenance" in the control plane operations type filter dropdown.·@nicoloboschi·691cb539

Bug Fixes

Make consolidation more reliable by retrying indefinitely with backoff and preventing duplicate work per bank.·@nicoloboschi·0d2ba56f
Fixed memory corruption when retaining large documents that exceeded a single LLM call — some extracted memories could be lost. Re-retain affected documents after upgrading.·@nicoloboschi·74525cc0
Improve Codex OAuth embeddings stability by adding token refresh support and follow-up fixes.·@DK09876·ffa6fbf2
Fix recall recency scoring to anchor timestamps to the query time for more consistent results.·@Sanderhoff-alt·2de19e57

0.7.0

Features

Added support for ParadeDB pg_search as a Citus-compatible BM25 search backend.·@nicoloboschi·4cd260b6
Added tokenizer configuration for pg_search to better control how text is tokenized for search.·@Sanderhoff-alt·16f80769
Added a new embeddings provider: ZeroEntropy.·@flamerged·ec49175f
Added support for Codex OAuth-based embeddings authentication.·@haha0815·d802f914
Added an API endpoint to clear mental model content.·@nicoloboschi·7bd11bed
Added targeted consolidation by observation scope to focus consolidation on specific parts of memory.·@nicoloboschi·ac3ab2b5
Added support for the Alibaba Qwen3 rerank model.·@quicklyfast·b83bb87d
Added Right Agent integration.·@onsails·203ddfdd
Added an OpenCode Go LLM provider.·@tevenfeng·c29c76e3

Improvements

Added configurable BM25 language/tokenization and optional PGroonga backend for better multilingual search tuning.·@nicoloboschi·cb04cb79
Added an environment variable to enable API access logging (Uvicorn access logs).·@nicoloboschi·25387083
Control plane now exposes a UI action to clear the mental model.·@nicoloboschi·6e9b741b
Improved observation consolidation quality, including better handling of temporal reasoning.·@nicoloboschi·d1ef9da9
Control plane now supports internationalization with 8 locales.·@nicoloboschi·486c3a8b

Bug Fixes

Fixed stale outgoing links after deletions by recomputing links asynchronously.·@nicoloboschi·cc3ba4a3
Fixed Codex configuration to ignore inherited base URLs and automatically refresh OAuth access tokens.·@benfrank241·795c081d
Stopped sending the temperature parameter to Anthropic requests to avoid API incompatibilities.·@nicoloboschi·dabbf9ff
Prevented duplicate webhook batch deliveries for retain events.·@xmh1011·6348f424
Improved reliability for large retain requests by splitting oversized single items in batch retains.·@nicoloboschi·c3b2b154
Fixed Ollama support by adding an Ollama Cloud provider and correcting authentication for cloud endpoints.·@nicoloboschi·cb037290
Ensured disabled tools are not included in the agent system prompt.·@nicoloboschi·2582b45a
Avoided retries when an embeddings provider returns invalid vector dimensions.·@ai-ag2026·e1e1a5e0
Prevented prompt formatting errors by escaping literal braces in all user-supplied prompt fields.·@nicoloboschi·67ae2a41
Fixed issues with mental-model baseline refresh and capped history size to prevent database JSON overflow.·@xmh1011·44b34c89
Enabled gzip handling to keep graph responses parseable when compressed.·@nicoloboschi·31d1e172
Added per-operation LLM concurrency caps to prevent overload and improve stability under parallel workloads.·@nicoloboschi·daf2348b
Preserved tag group/triggers correctly when updating tags.·@xmh1011·9c161e4e
Improved reranking robustness by detecting pre-normalized scores and applying rank-based normalization when needed.·@xuli500177·dcf5588e
Fixed entity resolution so user-defined label entities are not fuzzy-matched incorrectly.·@nicoloboschi·46dd2dfd
Improved worker resilience by handling stale pending schema routines.·@xmh1011·592f01bb
Fixed embedded UI startup on Windows by resolving the correct npx path before launch.·@tuancookiez-hub·2e5186a6
Improved control-plane session security by verifying signed session cookies instead of just checking presence.·@nicoloboschi·878ef957

0.6.2

Features

Added a configurable MCP request timeout to the Claude Code integration so long-running recall/reflect calls no longer time out at the default.·@rsaulo·55ef7067

Improvements

Renamed the Agent SDK agent_knowledge_recall max_results parameter to max_tokens and raised the default from 10 to 1024 so default-config users get useful recall results.·@r266-tech·d9dd1499
Applied the same max_results→max_tokens rename to the Claude Code MCP plugin so both integrations stay in sync.·@offendingcommit·909a4fd4

Security

Bumped vulnerable dependencies across the stack to address GitHub-reported CVEs — including litellm to ≥1.83.14, urllib3 to 2.7.0, and additional high/critical pip and npm packages.·@dcbouius·fd05bdab

Bug Fixes

Fixed the CLI “memory retain” timestamp/Event Date option so the provided timestamp is actually sent to the API.·@benfrank241·51ea9aa2
Repaired a database migration issue affecting mental model subtype data at the current schema head.·@benfrank241·debbd919
Improved migration robustness by handling transient database OID errors during the embedding-dimension migration.·@dcbouius·9dfbfb4b
Fixed Claude Code knowledge get_page calls to request detail=content and avoid tool-result spillover so full page content returns reliably.·@cdbartholomew·b2a693ab
Applied the same agent_knowledge_get_page detail=content fix to the Agent SDK so SDK callers also receive full page content.·@r266-tech·b593d40f
Aligned the Paperclip integration with Paperclip's actual event payload shape, restoring correct fact extraction from incoming events.·@amirhmoradi·be908d5b

0.6.1

Features

Added AlloyDB ScaNN vector index support, providing a high-performance vector index option for AlloyDB deployments·@can1357·e4422a9b
Added map-type entity labels so entity extraction can produce structured groups (e.g. address fields) instead of flat strings·@nicoloboschi·2b725bc4
Added z.ai (智谱) as a first-class LLM provider with a free-tier-friendly default model·@Burgunthy·4c75cd9e
Added a litellmrouter provider that lets the API call multiple LLM endpoints with automatic fallback chains·@nicoloboschi·98f33cbf
Added an optional read-only database backend for recall queries, letting recall traffic go to a replica while writes stay on the primary·@cdbartholomew·cf9b1f59
Added optional access-key login to the Control Plane for protecting the admin UI without a full SSO setup·@ariel-ai-bot·c0ff87ea
Added a --strategy flag to memory retain-files in the CLI for selecting how files are split before retention·@nicoloboschi·375747f5
Bank dropdown in the Control Plane now shows per-bank memory counts to make bank selection easier·@nicoloboschi·5a0cb4a5

Improvements

Worker progress-stats fanout is now scoped to schemas with pending work, eliminating wasted polling on idle tenants·@nicoloboschi·39d31ad2

Bug Fixes

Fixed fact extraction reliability by removing multiplicative retry layers that could turn one transient LLM failure into many·@nicoloboschi·a22e8bdd
Fixed reflect to read document metadata from the original retain parameters so reflect outputs include the right document context·@nepenth·f2a2f9fe
Failed batch_retain operations now propagate the child error_message to the parent, so failures surface a meaningful reason instead of a generic error·@cdbartholomew·7b82d05b
Fixed macOS daemon mode by replacing os.fork() with subprocess.Popen, restoring compatibility with PyTorch MPS·@nicoloboschi·8d77976a
Hardened Control Plane access-key authentication and improved login UX (clearer errors, redirect handling)·@nicoloboschi·b628716f
Reduced retain memory pressure by clearing content references after use, helping long-running workers avoid OOM on large batches·@nicoloboschi·c9145805
Fixed Docker image to chmod 755 /home/hindsight so containers run cleanly with --user UID:GID overrides·@nicoloboschi·a5cef602
Hard-pinned meta packages (hindsight-api, hindsight-all, hindsight-all-slim) to matching hindsight-api-slim, preventing stale slim installs after upgrade·@nicoloboschi·a86d5381
Worker now probes pg_proc before calling the optional schemas_with_pending_work() function, so older databases without it don't crash the worker·@nicoloboschi·a1c1b7de
TypeScript client now derives CLIENT_VERSION via tsup define so the published package version is always reported correctly·@nicoloboschi·ab8cc3e6

0.6.0

Features

Added Dify integration providing Hindsight memory tools inside Dify workflows.·@benfrank241·bc23750b
Added n8n community node package for using Hindsight memory in n8n automations.·@benfrank241·c1eaf711
Added SmolAgents integration providing Hindsight memory tools for SmolAgents-based agents.·@benfrank241·8314de5e
Added AWS Bedrock AgentCore Runtime integration (hindsight-agentcore).·@benfrank241·c91696f5
Added an Oracle Database backend option for enterprise storage.·@DK09876·50f559c9
Added the ability to cancel long-running async operations.·@nicoloboschi·7f30dcc7
Typescript client now supports AbortSignal across all methods for request cancellation.·@harryplusplus·8367930c

Improvements

Python client now exposes retain_async via retain()/aretain() for async memory retention flows.·@harryplusplus·daae8223
Anthropic provider now supports environment-configurable max retries and default headers.·@TuftyBruno·fa4bf700
Memories timeseries stats can now be grouped by different time fields (time_field toggle).·@aliu-ronin·cf1a97ab
Improved database performance for observation reads by reducing overhead and making reads backend-aware.·@nicoloboschi·d8ec2d7f

Bug Fixes

Fixed search ranking by correcting BM25 score direction for the vchord backend.·@liling·b322b0c5
Recall now correctly inherits observation entities through source_memory_ids, improving entity continuity across related memories.·@youchi1·8507095a
Entity co-occurrence timestamps now use the event date instead of the ingestion time, improving timeline accuracy.·@aliu-ronin·fc624cbf
Daemon mode now respects configured host settings (CLI --host and HINDSIGHT_API_HOST).·@nicoloboschi·3d3aa76b
Webhooks are now processed tenant-aware to prevent cross-tenant routing issues.·@cdbartholomew·b9069c28
API per-document graph and counts now include observations, improving completeness of document-level views.·@youchi1·7cc2daf4
Improved compatibility with OpenAI-compatible providers by hardening JSON response parsing and handling null content.·@voarsh2·bc14e5c4
Fixed Windows worker startup by handling platforms that don't implement add_signal_handler.·@nicoloboschi·08b56fdc
Retain/batch retain reliability improved by fixing transactional atomicity, recovery checkpoint scoping, and a cascade deadlock risk.·@cdbartholomew·f4ca3038
Typescript client now exposes previously missing recall/reflect parameters.·@nicoloboschi·641b3912
MCP recall tool now exposes tag_groups so recalls can be filtered/grouped by tag groups.·@nicoloboschi·b948b574
Embed/daemon startup reliability improved (script path resolution, correct extras when spawning API, and safer subprocess output handling).·@nicoloboschi·0f15f76a
Configuration logging now redacts database URLs to avoid leaking credentials in logs.·@xmh1011·2bada2db
Fixed support for OpenAI-compatible embedding dimensions by allowing provider-specific parameters.·@zwcf5200·324b4b0a

0.5.6

Bug Fixes

Reverted the JSON schema simplification introduced in 0.5.5 (5b1c3486). The change inlined $ref/$defs into a large blob that caused some LLMs (notably GPT-4o-mini in soft-enforcement mode) to echo the schema structure instead of producing valid responses, resulting in 0 facts extracted during retain. If you are on 0.5.5, upgrade to 0.5.6 immediately.

0.5.5

warning

0.5.5 contains a regression that can cause 0 facts to be extracted during retain. The JSON schema simplification change (5b1c3486) was reverted in 0.5.6. Please upgrade to 0.5.6.

Features

Added a new LlamaParse file parser, using the LlamaIndex Cloud parsing API to convert documents (PDF, DOCX, etc.) to markdown before retain.·@nicoloboschi·91106f30
Added a new Mental Models list view and tag filtering from mental models.·@nicoloboschi·8fbe85f0
Added Pipecat voice AI pipeline integration so voice agents can use Hindsight long-term memory.·@benfrank241·f7cc9ad6
Added an option to force IPv4 for Gemini embeddings to improve compatibility in restricted network environments.·@connorblack·6fb8ac97
Added support to control Cohere embedding output dimensions via configuration.·@nicoloboschi·a7514e18
Added a new Document Chunks API plus document reprocessing and a richer document detail experience in the UI.·8eb6e0e5b
Added an exclude_parents filter for list operations to better control what results are returned.·@nicoloboschi·8f6e0e5b
Operation responses now include retry information (retry count and next retry time).·@cdbartholomew·45f47a91
Retain results now include processed content token counts for better usage/throughput visibility.·@cdbartholomew·9c9d7917

Improvements

Added full Windows support for the embedded runtime and improved how the local API is discovered/launched.·@nicoloboschi·4ba54d8c
Operation status reporting now includes processing and cancelled states across the API and UI.·@nicoloboschi·80982da5
Workers can reserve per-operation slots when claiming tasks, improving fairness and throughput under load.·@nicoloboschi·c81e62ae
PostgreSQL connections now support a configurable statement timeout to prevent runaway queries.·@nicoloboschi·bdb3a55d
~~Improved LLM interoperability by simplifying JSON schemas and avoiding problematic tool-choice defaults.~~ Reverted in 0.5.6 — caused 0 facts extracted with some LLMs.·@nicoloboschi·5b1c3486
Worker scheduling no longer allows child tasks to block parent execution.·@nicoloboschi·a49d19cd

Bug Fixes

Fixed DeepSeek compatibility issues (including tool-calling quirks) and added it as a first-class LLM provider.·@nicoloboschi·461b00d4
Fixed Amazon Bedrock Converse compatibility by adjusting how causal relations are represented.·@nicoloboschi·4bc772d8
GET /banks/{bank_id}/profile no longer creates a bank as a side effect.·@cdbartholomew·99a89789
Reduced memory fan-out during consolidation recall to prevent excessive retrieval and improve stability.·@nicoloboschi·4ba2fffe
External PostgreSQL connection strings now accept asyncpg-style URLs.·@nicoloboschi·db7f4921
Timeseries stats buckets now return timezone-aware ISO timestamps.·@aliu-ronin·cd1ab497
Fixed conversation storage when using streaming LLM responses via LiteLLM.·@DK09876·ac5181f5
Removed the MCP delete_memory tool to close an authorization-bypass vulnerability.·@nicoloboschi·90674aef
First-time UI launch no longer requires manual confirmation to install the control plane.·@bwjoke·33aacf5c

0.5.4

Features

Delta mental model refresh now scopes recall to memories created since the last refresh, making updates faster and more accurate.·@nicoloboschi·e90cfa4a
OpenAI-compatible embedding providers now support configurable batch sizes for better throughput.·@r266-tech·30700de6
Embedded daemon now includes a liveness check that auto-recovers from crashes.·@nicoloboschi·59f9a2bf
Disable daemon idle timeout by default so embedded instances stay alive between requests.·@nicoloboschi·f5dfe59b
Add {user_id} template variable for retainTags in the Claude Code integration.·@soichisumi·9181c9a2
New decommission-workers and worker-status admin CLI commands for managing worker fleets.·@nicoloboschi·c8b898bd

Bug Fixes

Fix duplicate memory units caused by chunk index scrambling during concurrent upserts.·@nicoloboschi·511ca723
Prevent directive content from leaking into reflect answers on empty banks.·@nicoloboschi·3d877b05
Honor the reflect_mission identity framing in the prompt builder so agent personality works correctly.·@nicoloboschi·a3b0d265
Allow reflect-specific LLM configuration when the default LLM provider is disabled.·@zwcf5200·afd00c03
Preserve document created_at timestamp across upsert and add UI edit flow for documents.·@nicoloboschi·10785666
Fix ensure_ascii=False in json.dumps for LLM prompts so non-Latin text is preserved.·@harryplusplus·d05b49a2
Route update_bank through the config resolver with generic config_updates.·@nicoloboschi·abbd3619
Workers now scan for active schemas before claiming tasks, preventing work on decommissioned tenants.·@cdbartholomew·7126bf8a
Pass DeferOperation through MemoryEngine.execute_task so extensions can requeue work.·@cdbartholomew·858f0b3a
Downgrade LLM verify_connection failure to a warning instead of crashing on startup.·@nicoloboschi·9901aa1e
Fix items_count in list_operations response to populate from result_metadata.·@nicoloboschi·41710ba1
Align AI SDK ReflectBasedOn types with the OpenAPI spec.·@nicoloboschi·3d6b3805
Fix database migration path by merging divergent Alembic heads for v0.5.3.·@grimmjoww·487e2a5e
Lower OpenCode retainEveryNTurns default from 10 to 3 for more frequent memory saves.·@DK09876·902704df

0.5.3

Features

Add a setting to limit how many memories can be consolidated per round, helping control consolidation workload and cost.·@nicoloboschi·ca561aca
Add integration with the OpenAI Agents SDK.·@DK09876·b8da88c8
Improve mental model refresh and updates with structured operations and cleaner observation handling on upsert.·@nicoloboschi·8b80959b
Allow extensions to requeue work by throwing a DeferOperation exception from worker operations.·@nicoloboschi·f8904797
Make recall budget mapping configurable per memory bank.·@nicoloboschi·576c44d2
Control plane now shows failed consolidation counts with a drilldown to investigate issues.·@nicoloboschi·e1e5f36c
Add mental-model staleness signals and a refreshed UI/experience for reviewing model history and snapshots.·@nicoloboschi·654e4c0c
Replace the embedded Paperclip library with the Paperclip plugin for more flexible integrations.·@benfrank241·c571fac7
JSON logs can now include tenant information and support a configurable allowlist for what gets logged.·@nicoloboschi·3bedc1ce
CLI now supports named connection profiles (via -p/--profile) for easier switching between environments.·@nicoloboschi·70d60e96

Improvements

Reduce the default number of retries for LLM calls to fail faster when providers are erroring or unavailable.·@nicoloboschi·b52b483c
Make reranker failures easier to diagnose and add a configurable timeout for the TEI reranker.·@octo-patch·69383af8

Bug Fixes

Fix crashes when using Jina MLX on Metal GPUs by serializing GPU operations.·@lkttle·2e74a324
Fix database migration path so upgrades from v0.4.22 to v0.5.x work correctly.·@nicoloboschi·5437cc02
Prevent orphaned observations if a source memory is deleted during consolidation.·@nicoloboschi·f9042e37
Fix Ollama requests by explicitly disabling "think" mode in the native call payload.·@karl-8888·7d4fd1aa
Fix file retain uploads and prevent orphaned retained files.·@christerence·9e30ae25
Fix file retain API to correctly accept and map a provided "timestamp" field.·@christerence·13f3052e
Improve fairness across tenants when workers claim tasks, reducing the chance of noisy tenants starving others.·@cdbartholomew·a5e53721
Ensure the mental model max_tokens setting is respected during refresh/reflect operations.·@nicoloboschi·568e3c30
Fix control-plane links by properly encoding bank IDs in URLs end-to-end.·@nicoloboschi·cbaec36f
Make task submission idempotent when a payload is already set, preventing duplicate/failed submissions.·@nicoloboschi·088dfecb

0.5.2

Features

Added a co-occurrence graph view for exploring entity relationships in the control plane.·@nicoloboschi·f64c5d20
Added recall controls to the mental model trigger API/CLI so you can tune what gets recalled during runs.·@nicoloboschi·f2fc8f9f
Async operations now expose task payload details and associated document IDs for better observability and debugging.·@nicoloboschi·870bf4a3

Improvements

Revamped the control plane bank statistics view for clearer insights.·@nicoloboschi·34365c32
Clients now send an identifying User-Agent header on all HTTP requests for easier server-side diagnostics.·@nicoloboschi·9372462e

Bug Fixes

Fixed consolidation retry budget handling so retries are correctly applied at the LLM call site.·@r266-tech·dee58139
Fixed a crash during retain when embeddings and extracted facts counts didn’t match.·@nicoloboschi·dbd1d1a7
Improved embedded mode cleanup stability by adding a timeout when acquiring the cleanup lock (prevents hangs).·@r266-tech·6b5aa3af
OpenClaw plugins now reliably register agent hooks on every entry invocation.·@nicoloboschi·1be5ff33
TypeScript SDK now re-exports BankTemplate types from the package root for simpler imports.·@mrkhachaturov·581bbf3f
Bank template configuration validation was aligned with configurable fields to prevent invalid/ignored settings.·@mrkhachaturov·099f4c92

0.5.1

Breaking Changes

OpenClaw now reads configuration from plugin config instead of environment variables. (e22ae05f)

Features

Added SiliconFlow as a supported reranker provider. (d0b2ab9a)
Added an interactive OpenClaw setup wizard with Cloud / API / Embedded modes. (87322396)
Added a config-aware CLI to backfill OpenClaw history. (72fd3d59)
Added OpenClaw session pattern filtering to ignore or treat sessions as stateless. (5a61ac50)
Added a Cloudflare OAuth proxy integration option for self-hosted Hindsight. (aad07a14)
Expanded the CLI to cover all OpenAPI endpoints and request-body parameters. (c05c491d)
Added a default bank template environment variable (HINDSIGHT_API_DEFAULT_BANK_TEMPLATE). (fc941d5c)
Added a daemon lifecycle package (@vectorize-io/hindsight-all) to simplify running the all-in-one daemon. (576016f5)
Added recallTags and recallTagsMatch configuration options to control which tagged memories are recalled. (b57e337f)

Improvements

Improved OpenClaw reliability with more resilient startup behavior and richer retain metadata. (1f1716bd)

Bug Fixes

OpenClaw setup wizard now prompts for the token value (not the env var name). (9679d813)
Fixed embedded mode daemon start/stop race that could terminate healthy daemons. (e5724fcb)
Fixed reranker initialization issues to show real import errors and avoid a Transformers 5.x race in jina-mlx. (f82f58fa)
Fixed worker consolidation slot accounting to respect the configured maximum concurrency. (2d74007d)
Improved CLI API error output by including the HTTP response body. (93300b91)
Fixed CLI memory listing showing "[UNKNOWN]" for fact types. (2635bbb4)
Fixed recall ranking so RRF ordering is preserved when the reranker is configured as a passthrough. (4f9cf15c)
Fixed retain chunk insertion to be idempotent and avoid repeated retries on integrity errors. (2d95f78b)
Fixed retain ANN seed temp table creation to run inside a transaction for better reliability. (3fc87e76)
Fixed LLM requests to use the correct max token parameter for reasoning models and Azure OpenAI. (7b2263ba)

0.5.0

Breaking Changes

Removed BFS and MPFP graph retrieval strategies. LinkExpansionRetriever is now the sole graph retrieval algorithm, offering simpler, faster, and more accurate results. (ea834bc7)
Dropped the hindsight-hermes integration package. (cf0537ba)

Features

Built-in llama.cpp LLM provider for fully local inference without external API calls. (f74b577e)
Retain update_mode='append' for concatenating new content onto an existing document instead of replacing it. (3c633e5e)
OpenRouter support for LLM, embeddings, and reranking. (e5944b63)
Bank template import/export with Template Hub — export a bank's configuration, mental models, and directives as a reusable manifest, then import into other banks. (30a319a6)
Constellation view in the Control Plane — interactive, zoomable canvas visualization of entity relationship graphs with heat-gradient coloring and dark mode support. (36783df3)
Added detail parameter to list/get mental model endpoints for controlling response verbosity. (8d1bfbbd)
Added AutoGen integration (hindsight-autogen) for persistent long-term memory in AutoGen agents. (a757765a)
Added Paperclip integration (@vectorize-io/hindsight-paperclip) with Express middleware and process adapter modes for stateless agent memory. (81441ee9)
Added OpenCode persistent memory plugin for the OpenCode editor. (e1c6220f)
OpenClaw JSONL-backed retain queue for external API resilience — buffers retain calls locally when the API is unreachable. (087545cc)
OpenClaw now supports bankId for static bank configurations. (0e81d1a2)
Added Google embeddings and reranker provider support. (07de798c)
Added persistent volume support in Helm chart for local model cache. (cefa7554)
MCP server now includes a sync_retain tool and validates UUID inputs. (48185a4b)
Recall combined scoring now includes proof_count boost for better ranking. (26794aab)

Improvements

3-phase retain pipeline restructures memory ingestion into pre-resolve, insert, and post-link phases, dramatically improving throughput under concurrent load by removing slow reads from write transactions. (914ba796)
Recall entity graph expansion now caps per-entity fanout and includes a timeout fallback, preventing slow queries on banks with high-fanout entities. (57f15445)
Fact serialization in think-prompt now includes occurred_end and mentioned_at for richer temporal context. (37348c85)
Consolidation observation quality improved with structured processing rules. (6f173b10)

Bug Fixes

LiteLLM SDK embeddings encoding_format is now configurable instead of hardcoded. (cece2c90)
Fixed out-of-range content_index crash in recall result mapping. (9790d904)
Experience fact types are now preserved correctly during normalization. (9cfdd464)
Clear memories endpoint no longer deletes the bank profile. (26a64cc0)
Embedding daemon clears stale processes on the port before starting. (7d6c570a)
Per-bank vector index migration now respects vector extension configuration. (4fd7c5d1)
Timeline group sort uses numeric date comparison instead of locale string comparison. (f3f2c6b0)
Resolved 25 test regressions from the streaming retain pipeline. (7415ebff)
MCP server now auto-coerces string-encoded JSON in tool arguments. (443c94c8)
Entity labels structure is now validated on PATCH to prevent invalid configurations. (7e23f8e1)
Fixed bank_id metric label to be opt-in, preventing OTel memory leak. (cf4bd598)
Fixed max_tokens handling for OpenAI-compatible endpoints with custom base URLs. (cd99eef4)
Fixed event_date AttributeError when date is None in fact extraction. (6cb309f7)
Query analyzer now handles dateparser internal crashes gracefully. (e0e65c44)
Embedding profile .env overwrite skipped when config has no Hindsight keys. (9e2890ba)
Windows compatibility fix for hindsight-embed. (f9fe6953)
Addressed critical and high severity security vulnerabilities in dependencies. (ee4510a7)

0.4.22

Features

API now supports passing custom LLM request parameters via the HINDSIGHT_API_LLM_EXTRA_BODY configuration. (ecaa1ad1)
Document metadata is now exposed through the API and control plane. (627ec5d5)
Added a /code-review skill for automated code quality checks against project standards. (bdb33c58)
ZeroEntropy reranker now supports a configurable base URL. (a915584e)
Codex can now retain structured tool calls from rollout files. (3461398b)

Improvements

Embeddings via the LiteLLM SDK can now optionally specify output dimensions. (f841bcb9)
API responses now include an X-Ignored-Params header to warn when unknown request parameters were ignored. (cef42d81)
OpenClaw CLI startup is faster by deferring heavy initialization until the service starts. (41025c3b)

Bug Fixes

Mental model triggers now support the full config schema, including tag matching and tag group filters. (2c32ffad)
Cohere reranking via Azure endpoints now works reliably (avoids 404 errors). (84985ee9)
Claude Code provider no longer defers to built-in tools, preventing MCP tool handling issues. (fa82efc8)
Recall endpoint now returns metadata correctly instead of dropping it from the response. (4768bf39)
Gemini 3.1+ tool calls now read thought signatures correctly. (1b5c262a)
First-person agent memories are now correctly classified as "experience" facts. (00961156)
Codex upgrades now preserve and merge new settings instead of skipping them. (b104bad0)
LlamaIndex integration fixes improve document ID handling, memory API behavior, and ReAct tracing. (d93dfea8)

0.4.21

Features

Added audit logging for feature usage tracking, including request duration in audit entries. (083295dc)
Added Hindsight memory integration for the OpenAI Codex CLI. (0b17a67c)
Added an MCP hook to filter tool visibility per user. (f8285b7b)
Added a per-bank limit setting to cap the number of observations stored per scope. (b32767ca)
Added native Windows support so Hindsight can run without Docker. (c5700ff5)
Added a 'none' LLM provider to support chunk-only storage without LLM calls. (9e5a066d)
Added a setup command/skill to register hooks more reliably. (22ca6a8d)
Hermes now supports file-based configuration. (0ff36548)
Added a LiteLLM-based provider to support Bedrock and many additional LLM providers. (db70fdbe)
Added support for Strands Agents SDK integration with Hindsight memory tools. (7fe773c0)
Added LlamaIndex integration. (2d787c4f)
Added AG2 framework integration. (73123870)
Added support for Ark and Volcano LLM providers. (417fac61)
Retain now supports delta mode to skip LLM processing for unchanged chunks on upsert. (fd88c0ef)
Claude Code integration can now retain full sessions with document upsert and configurable tags, and records tool calls as structured JSON. (2d31b67d)
MCP retain tool now supports selecting a retain strategy via a parameter. (4285e944)

Improvements

OpenClaw logging is now configurable and can emit structured output. (d441ab81)
Made inclusion of source facts in search observations configurable. (5095d5e3)
Integrations no longer use hardcoded default models, relying on configured defaults instead. (58e68f3e)

Bug Fixes

Improved MCP server compatibility by handling Claude Code GET probes and allowing stateless HTTP mode to be configured. (d8050387)
Per-bank vector index creation now respects the configured vector extension setting. (6488c9bc)
Verbose retain extraction now correctly includes the retain mission context. (d2965e64)
Codex integration no longer crashes on startup when the API quota is exhausted (HTTP 429). (111e8c70)
OpenAI embeddings client now correctly parses query parameters included in base_url. (a209ef1a)
Fixed tool_choice handling for Codex/Claude Code when forcing specific tool calls. (585ac76f)
OpenClaw auto-recall now supports a configurable timeout to prevent hangs. (cd4d449f)
Fixed control plane UI issues affecting recall and data viewing. (6bb83f46)
Recall responses now include associated metadata. (0bcbf849)
Python client update_bank_config() now exposes all configurable fields. (7c18723f)
API OpenAPI schema now correctly includes Pydantic v2 ValidationError fields. (939cb40a)
JSON-string tags are now coerced to lists for MemoryItem and MCP tools to prevent tagging errors. (c5273f5f)

0.4.20

Features

Add a one-command setup CLI package for the NemoClaw integration. (d284de28)
Add a LangGraph integration for using Hindsight memory within LangGraph agents. (b4320254)
Add reflect filters to exclude specific fact types and mental model content during reflection. (ea662d06)
Introduce independent versioning for integrations so they can be released separately from the core server. (31f1c53c)
Add a Claude Code integration plugin. (f4390bdc)

Improvements

Add a wall-clock timeout to reflect operations so they don’t run indefinitely. (8ce06e3e)
Provide richer context when validating operations via the OperationValidator extension. (2eb1019d)
Make the hindsight-api package runnable directly via uvx by adding script entry points. (97f7a365)
Support passing query parameters during OpenAI-compatible client initialization for broader provider compatibility. (20e17f28)
Upgrade the default MiniMax model from M2.5 to M2.7. (1f1462a5)

Bug Fixes

Prevent context overflow during observation search by disabling source facts in results. (8e2e2d5b)
Fix Claude Code integration session startup by pre-starting the daemon in the background. (26944e25)
Fix Claude Code integration installation and configuration experience so setup is more reliable. (35b2cbb6)
Fix a memory leak in entity resolution that could grow over time under load. (e6333719)
Avoid crashes and retain failures when the Postgres pg_trgm extension is unavailable by handling detection/fallback correctly. (365fa3ce)
Strip Markdown code fences from model outputs across all LLM providers for more consistent parsing. (2f2db2a6)
Return a clear 400 error for empty recall queries and fix a SQL parameterization issue. (5cdc714a)
Ensure file retain requests include authentication headers so uploads work in authenticated deployments. (78aa7c53)
Fix MCP tool calls when MCP_AUTH_TOKEN and TENANT_API_KEY differ. (8364b9c5)
Allow claude-agent-sdk to install correctly on Linux/Docker environments. (3f31cbf5)
In LiteLLM mode, fall back to the last user message when no explicit hindsight query is provided. (5e8952c5)
Fix non-atomic async operation creation to prevent inconsistent operation records. (94cf89b5)
Prevent orphaned parent operations when a batch retain child fails unexpectedly. (43942455)
Fix failures for non-ASCII entity names by ensuring entity IDs are set correctly. (438ce98b)
Correctly store LLM facts labeled as "assistant" as "experience" in the database. (446c75f3)

0.4.19

Features

TypeScript client now works in Deno environments. (72c25c97)
Added Agno integration to use Hindsight as a memory toolkit. (8c378b98)
Added Hermes Agent integration (hindsight-hermes) for persistent memory. (ef90842f)
Expanded retain behavior with new verbatim and chunks extraction modes and named retain strategies. (e4f8a157)

Improvements

Improved local reranker performance/efficiency with FP16 and bucketed batching, plus compatibility with Transformers 5.x. (e7da7d0e)

Bug Fixes

Prevented silent memory loss when consolidation fails (failed consolidations are tracked and can be recovered). (28dac7c7)
Fixed Docker control-plane startup to respect the configured control-plane hostname. (8a64dc8d)
Database cleanup migration now removes orphaned observation memory units to avoid inconsistent memory state. (f09ad9de)
Deleting a document now also deletes linked memory units to prevent leftover/stale memory entries. (f27bd953)
Fixed MCP middleware to send an Accept header, preventing 406 response errors in some setups. (836fd81e)
Improved compatibility with Gemini tool-calling by preserving thought signature metadata to avoid failures on gemini-3.1-flash-lite-preview. (21f9f46c)

0.4.18

Features

Add compound tag filtering using tag groups. (5de793ee)
Publish new slim Python packages (hindsight-api-slim and hindsight-all-slim) for smaller installs. (15ea23d5)
Add MiniMax as a supported LLM provider. (2344484f)
Add Jina MLX reranker provider optimized for Apple Silicon. (1caf5ec9)

Improvements

Allow configuring maximum recall query tokens via an environment variable. (66dedb8d)
Improve retrieval performance by switching to per-bank HNSW indexes. (43b3efc4)

Bug Fixes

Prevent reranking failures by truncating long documents that exceed LiteLLM reranker context limits. (eeb938fc)
Ensure recalled memories are injected as system context for OpenClaw. (b17f338e)
Ensure embedded profiles are registered in CLI metadata when the daemon starts. (06b0f74a)
Cancel in-flight async operations when a bank is deleted to avoid dangling work. (0560f626)

0.4.17

Features

Added a manual retry option for failed asynchronous operations. (dcaacbe4)
You can now change/update tags on an existing document. (1b4ad7f4)
Added history tracking and a diff view for mental model changes. (e2baca8b)
Added observation history tracking with a UI diff view to review changes over time. (576473b6)
File uploads can now choose a parser per request, with configurable fallback chains. (99220d05)
Added an extension hook that runs after file-to-Markdown conversion completes. (1d17dea2)

Improvements

Operations view now supports filtering by operation type and has more reliable auto-refresh behavior. (f7a60f89)
Added token limits for “source facts” used during consolidation and recall to better control context usage. (5d05962d)
Improved bank selector usability by truncating very long bank names in the dropdown. (1e40cd22)

Bug Fixes

Fixed webhook schema issues affecting multi-tenant retain webhooks. (32a4882a)
Fixed file ingestion failures by stripping null bytes from parsed file content before retaining. (cd3a6a22)
Fixed tool selection handling for OpenAI-compatible providers when using named tool_choice. (1cdfb7c2)
Improved consolidation behavior to prioritize a bank’s mission over an ephemeral-state heuristic. (00ccf0b2)
Fixed database migrations to correctly handle mental model embedding dimension changes. (7accac94)
Fixed file upload failures caused by an Iris parser httpx read timeout. (fa3501d4)
Improved reliability of running migrations by serializing Alembic upgrades within the process. (f88b50a4)
Fixed Google Cloud Storage authentication when using Workload Identity Federation credentials. (d2504ac5)
Fixed the bank selector to refresh the bank list when the dropdown is opened. (0ad8c2d0)

0.4.16

Features

Added Webhooks with consolidation.completed and retain.completed events. (abbf874d)

Improvements

Improved OpenClaw recall/retention controls. (d425e93c)
Improved search/reranking quality by switching combined scoring to multiplicative boosts. (aa8e5475)
Improved performance of observation recall by 40x on large banks. (ad2cf72a)
Improved server shutdown behavior by capping graceful shutdown time and allowing a forced kill on a second Ctrl+C. (4c058b4b)

Bug Fixes

Fixed an async deadlock risk by running database schema migrations in a background thread during startup. (e0a2ac63)
Fixed webhook delivery/outbox processing so transactions don’t silently roll back due to using the wrong database schema name. (75b95106)
Fixed observation results to correctly resolve and return related chunks using source_memory_ids. (cb6d1c46)
Fixed MCP bank-level tool filtering compatibility with FastMCP 3.x. (f17406fd)
Fixed crashes when an LLM returns invalid JSON across all retries (now handled cleanly instead of raising a TypeError). (66423b85)
Fixed observations without source dates to preserve missing (None) temporal fields instead of incorrectly populating them. (891c33b1)

0.4.15

Features

Added observation_scopes to control the granularity/visibility of observations. (55af4681)
List documents API now supports filtering by tags (and fixes the q parameter description). (1d70abfe)
Added PydanticAI integration for persistent agent memory. (cab5a40f)
Added richer entity label support (optional labels, free-form values, multi-value fields, and UI polish). (9b96becc)
Added support for timestamp="unset" so content can be retained without a date. (f903948a)
OpenClaw can now automatically retain the last n+2 turns every n turns (default n=10). (ad1660b3)
Added configurable Gemini/Vertex AI safety settings for LLM calls. (73ef99e7)
Added extension hooks to customize root routing and error headers. (e407f4bc)

Improvements

Improved recall performance by fetching all recall chunks in a single query. (61bf428b)
Improved recall/retain performance and scalability for large memory banks. (7942f181)

Bug Fixes

Fixed the TypeScript SDK to send null (not undefined) when includeEntities is false. (15f4b876)
Prevented reflect from failing with context_length_exceeded on large memory banks. (77defd96)
Fixed a consolidation deadlock caused by retrying after zombie processing tasks. (c2876490)
Fixed observations count in the control plane that always showed 0. (eaeaa1f2)
Fixed ZeroEntropy rerank endpoint URL and ensured the MCP retain async_processing parameter is handled correctly. (f6f1a7d8)
Fixed JSON serialization issues and logging-related exception propagation when using the claude_code LLM provider. (ecb833f4)
Added bank-scoped request validation to prevent cross-bank/invalid bank operations. (5270aa5a)

0.4.14

Features

Add Chat SDK integration to give chatbots persistent memory. (fed987f9)
Allow configuring which MCP tools are exposed per memory bank, and expand the MCP tool set with additional tools and parameters. (3ffec650)
Enable the bank configuration API by default. (4d030707)
Support filtering graph-based memory retrieval by tags. (0bb5ca4c)
Add batch observations consolidation to process multiple observations more efficiently. (0aa7c2b3)
Add OpenClaw options to toggle autoRecall and exclude specific providers. (3f9eb27c)
Add a ZeroEntropy reranker provider option. (17259675)

Improvements

Increase customization options for reflect, retain, and consolidation behavior. (2a322732)
Include source document metadata in fact extraction results. (87219b73)

Bug Fixes

Raise a clear error when embedding dimensions exceed pgvector HNSW limits (instead of failing later at runtime). (8cd65b98)
Fix multi-tenant schema isolation issues in storage and the bank config API. (b180b3ad)
Ensure LiteLLM embedding calls use the correct float encoding format to prevent embedding failures. (58f2de70)
Improve recall performance by reducing memory usage during retrieval. (9f0c031d)
Handle observation regeneration correctly when underlying memories are deleted. (ac9a94ad)
Fix reflect retrieval to correctly populate dependencies and enforce full hierarchical retrieval. (8b1a4658)
Fix OpenClaw health checks by passing the auth token to the health endpoint. (40b02645)

0.4.13

Features

Switched the default OpenAI LLM to gpt-4o-mini. (325b5cc1)
Observation recall now includes the source facts behind recalled observations. (5569d4ad)
Added CrewAI integration to enable persistent memory. (41db2960)

Bug Fixes

Fixed npx hindsight-control-plane failing to run. (0758827d)
Improved MCP compatibility by aligning the local MCP implementation with the server and removing the deprecated stateless parameter. (ea8163c5)
Fixed Docker startup failures when using named Docker volumes. (ac739487)
Prevented reranker crashes when an upstream provider returns an error. (58c4d657)
Improved accuracy of fact temporal ordering by reducing per-fact time offsets. (c3ef1555)
Client timeout settings are now properly respected. (dcaa9f14)
Fixed documents not being tracked when fact extraction returns zero facts. (f78278ea)

0.4.12

Features

Accept and ingest PDFs, images, and common Office documents as inputs. (224b7b74)
Add the Iris file parser for improved document parsing support. (7eafba66)
Add async Retain support via provider Batch APIs (e.g., OpenAI and Groq) for higher-throughput ingestion. (40d42c58)
Allow Recall to return chunks only (no memories) by setting max_tokens=0. (7dad9da0)
Add a Go client SDK for the Hindsight API. (2a47389f)
Add support for the pgvectorscale (DiskANN) vector index backend. (95c42204)
Add support for Azure pg_diskann vector indexing. (476726c2)

Improvements

Improve reliability of async batch Retain when ingesting large payloads. (aefb3fcf)
Improve AI SDK tooling to make it easier to work with Hindsight programmatically. (d06a0259)

Bug Fixes

Ensure document tags are preserved when using the async Retain flow. (b4b5c44a)
Fix OpenClaw ingestion failures for very large content (E2BIG). (6bad6673)
Harden OpenClaw behavior (safer shell usage, better HTTP mode handling, and more reliable initialization), including per-user banks support. (c4610130)
Improve Python client async API consistency and reduce connection drop issues via keepalive timeout fixes. (8114ef44)

0.4.11

Features

Added support for LiteLLM SDK as an embeddings and reranking provider. (e408b7e)
Expanded Postgres search support with additional text/vector extensions, including TimescaleDB pg_textsearch and vchord/pgvector options. (d871c30)
Added hierarchical configuration scopes (system, tenant, bank) for more flexible multi-tenant setup and overrides. (8d731f2)
Added reverse proxy/base-path support for running Hindsight behind a proxy. (93ddd41)
Added MCP tools to create, read, update, and delete mental models. (f641b30)
Added a "docs" skill for agents/tools to access documentation-oriented capabilities. (dd1e098)
Added an OpenClaw configuration option to skip recall/retain for specific providers. (fb7be3e)

Improvements

Improved LiteLLM gateway model configuration for more reliable provider/model selection. (7d95a00)
Exposed actual LLM token usage in retain results to improve cost/usage visibility. (83ca669)
Added user-initiated attribution to request context to improve async task and usage attribution. (90be7c6)
Added OpenTelemetry tracing for improved request traceability and observability. (69dec8e)
Helm chart: split TEI embedding and reranker into separate deployments for independent scaling and rollout. (43f9a8b)
Helm chart: added PodDisruptionBudgets and per-component affinity controls for more resilient scheduling. (9943957)

Bug Fixes

Fixed a recursion issue in memory retention that could cause failures or runaway memory usage. (4f11210)
Fixed Reflect API serialization/schema issues for "based_on" so reflections are returned and stored correctly. (f9a8a8e)
Improved MCP server compatibility by allowing extra tool arguments when appropriate and fixing bank ID resolution priority. (7ee229b)
Added missing trust_code environment configuration support. (60574ee)
Hardened the MCP server with fixes to routing/validation and more accurate usage metering. (e798979)
Fixed the slim Docker image to include tiktoken to prevent runtime tokenization errors. (6eec83b)
Fixed MCP operations not being tracked correctly for usage metering. (888b50d)
Helm chart: fixed GKE deployments overriding the configured HINDSIGHT_API_PORT. (03f47e2)

0.4.10

Features

Provided a slimmer Docker distribution to reduce image size and speed up pulls. (f648178)
Added Markdown support in Reflect and Mental Models content. (c4ef090)
Added built-in Supabase tenant extension for running Hindsight with Supabase-backed multi-tenancy. (e99ee0f)
Added TenantExtension authentication support to the MCP endpoint. (fedfb49)

Improvements

Improved MCP tool availability/routing based on the endpoint being used. (d90588b)

Bug Fixes

Stopped logging database usernames and passwords to prevent credential leaks in logs. (c568094)
Fixed OpenClaw sessions wiping memory on each new session. (981cf60)
Fixed hindsight-embed profiles not loading correctly. (0430588)
Fixed tagged directives so they correctly apply to tagged mental models. (278718d)
Fixed a cast error that could cause failures at runtime. (093ecff)

Other

Added a docker-compose example to simplify local deployment and testing. (5179d5f)

0.4.9

Features

New AI SDK integration. (7e339e1)
Add a Python SDK for running Hindsight in embedded mode (HindsightEmbedded). (d3302c9)
Add streaming support to the hindsight-litellm wrappers. (665877b)
Add OpenClaw support for connecting to an external Hindsight API and using dynamic per-channel memory banks. (6b34692)

Improvements

Improve the mental models experience in the control plane UI. (7097716)
Reduce noisy Hugging Face logging output. (34d9188)

Bug Fixes

Improve recall endpoint reliability by handling timeouts correctly and rejecting overly long queries. (dd621a6)
Improve /reflect behavior with Claude Code and Codex providers. (a43d208)
Fix OpenClaw shell argument escaping for more reliable command execution. (63e2964)

0.4.8

Features

Added profile support for hindsight-embed, enabling separate embedding configurations/workspaces. (6c7f057)
Added support for additional LLM backends, including OpenAI Codex and Claude Code. (539190b)

Improvements

Enhanced OpenClaw and hindsight-embed parameter/config options for easier configuration and better defaults. (749478d)
Added OpenClaw plugin configuration options to select LLM provider and model. (8564135)
Server now prints its version during startup to simplify debugging and support requests. (1499ce5)
Improved tracing/debuggability by propagating request context through asynchronous background tasks. (44d9125)
Added stronger validation and context for mental model create/refresh operations to prevent invalid requests. (35127d5)

Bug Fixes

Improved embedding CLI experience with richer logs and isolated profiles to avoid cross-contamination between runs. (794a743)
Operation validation now runs correctly in the worker process, preventing invalid background operations from slipping through. (96f0e54)
Fixed unreliable behavior when using a custom PostgreSQL schema. (3825506)

0.4.7

Features

Add extension hooks to validate and customize mental model operations. (9c3fda7)
Add support for using an external embedding API provider in OpenClaw plugin (with additional OpenClaw compatibility fixes). (4b57b82)

Improvements

Speed up container startup by preloading the tiktoken encoding during Docker image builds. (039944c)

Bug Fixes

Prevent PostgreSQL insert failures by stripping null bytes from text fields before saving. (ef9d3a1)
Fix worker schema selection so it uses the correct default database schema. (d788a55)
Honor an already-set HINDSIGHT_API_DATABASE_URL instead of overwriting it in the hindsight-embed workflow. (f0cb192)

0.4.6

Improvements

Improved OpenClaw configuration setup to make embedding integration easier to configure. (27498f9)

Bug Fixes

Fixed OpenClaw embedding version binding/versioning to prevent mismatches when using the embed integration. (1163b1f)

0.4.5

Bug Fixes

Fixed occasional failures when retaining memories asynchronously with timestamps. (cbb8fc6)

0.4.4

Bug Fixes

Fixed async “retain” operations failing when a timestamp is provided. (35f0984)
Corrected the OpenClaw daemon integration name to “openclaw” (previously “openclawd”). (b364bc3)

0.4.3

Features

Add Vertex AI as a supported LLM provider. (c2ac7d0, 49ae55a)
Add Bearer token authentication for MCP and propagate tenant authentication across MCP requests. (0da77ce)

Improvements

CLI: add a --wait flag for consolidate and a --date filter for listing documents. (ff20bf9)

Bug Fixes

Fix worker polling deadlocks to prevent background processing from stalling. (f4f86e3)
Improve reliability of Docker builds by retrying ML model downloads. (ecc590c)
Fix tenant authentication handling for internal background tasks and ensure the control-plane forwards required auth to the dataplane. (03bf13e)
Ensure tenant database migrations run at startup and workers use the correct tenant schema context. (657fe02)
Fix control-plane graph endpoint errors when upstream data is missing. (751f99a)

Other

Rename the default bot/user identity from "moltbot" to "openclaw". (728ce13)

0.4.2

Features

Added Clawdbot/Moltbot/OpenClaw integration. (12e9a3d)

Improvements

Added additional configuration options to control LLM retry behavior. (3f211f0)
Added real-time logs showing a detailed timing breakdown during consolidation runs. (8781c9f)

Bug Fixes

Fixed hindsight-embed crashing on macOS. (c16ccc2)

0.4.1

Features

Added support for using a non-default PostgreSQL schema by default. (2b72e1f)

Improvements

Improved memory consolidation performance (benchmarking and optimizations). (b43ef98)

Bug Fixes

Fixed the /version endpoint returning an incorrect version. (cfcc23c)
Fixed mental model search failing due to UUID type mismatch after text-ID migration. (94cc0a1)
Added safer PyTorch device detection to prevent crashes on some environments. (67c4788)
Fixed Python packages exposing an incorrect version value. (fccbdfe)

0.4.0

Observations, Mental Models, new Agentic Reflect and Directives, read the announcement.

Features

Added support for providing a custom prompt for memory extraction. (3172e99)
Expanded the LiteLLM integration with async retain/reflect support, cleaner API, and support for tags/mission (including passing API keys correctly). (1d4879a)
Added a new worker service to run background tasks at scale. (4c79240)
MCP retain now supports timestamps. (b378f68)
Added support for installing skills via npx add-skill. (ec22317)

Improvements

CLI retain-files now accepts more file types. (1eeced3)

Bug Fixes

Fixed a macOS crash in the embed daemon caused by an XPC connection issue. (e5fc6ee)
Fixed occasional extraction in the wrong language. (87d4a36)
Fixed PyTorch model initialization issues that could cause startup failures (meta tensor/init problems). (ddaa5f5)

Features

Add memory tags so you can label and filter memories during recall/reflect. (20c8f8b)
Allow choosing different AI providers/models per operation. (e6709d5)
Add Cohere support for embeddings and reranking. (4de0730)
Add configurable embedding dimensions and OpenAI embeddings support. (70de23e)
Support custom base URLs for OpenAI-style embeddings and Cohere endpoints. (fa53917)
Add LiteLLM gateway support for routing LLM/embedding requests. (d47c8a2)
Add multilingual content support to improve handling and retrieval across languages. (c65c6a9)
Add delete memory bank capability. (4b82d2d)
Add backup/restore tooling for memory banks. (67b273d)

Improvements

Add retention modes to control how memories are extracted and stored. (fb31a35)
Add offline (optional) database migrations to support restricted/air-gapped deployments. (233bd2e)
Add database connection configuration options for more flexible deployments. (33fac2c)
Load .env automatically on startup to simplify configuration. (c06d9b4)
Expose an operation ID from retain requests so async/background processing can be tracked. (1dacd0e)
Add per-request LLM token usage metrics for monitoring and cost tracking. (29a542d)
Add LLM call latency metrics for performance monitoring. (5e1f13e)
Include tenant in metrics labels for better multi-tenant observability. (1ffc2a4)
Add async processing option to MCP retain tool for background retention workflows. (37fc7fb)

Bug Fixes

Fix extension loading in multi-worker deployments so all workers load extensions correctly. (f5f3fca)
Improve recall performance by batching recall queries. (5991308)
Improve retrieval quality and stability for large memory banks (graph/MPFP retrieval fixes). (6232e69)
Fix entities list being limited to 100 entities. (26bf571)
Fix UI only showing the first 1000 memories. (67c1a42)
Fix duplicated causal relationships and improve token usage during processing. (49e233c)
Improve causal link detection accuracy. (2a00df0)
Make retain max completion tokens configurable to prevent truncation issues. (7715a51)
Fix Python SDK not sending the Authorization header, preventing authenticated requests. (39e3f7c)
Fix stats endpoint missing tenant authentication in multi-tenant setups. (d6ff191)
Fix embedding dimension handling for tenant schemas in multi-tenant databases. (6fe9314)
Fix Groq free-tier compatibility so requests work correctly. (d899d18)
Fix security vulnerability (qs / CVE-2025-15284). (b3becb6)
Restore MCP tools for listing and creating memory banks. (9fd5679)

0.2.0

Features

Add additional model provider support, including Anthropic Claude and LM Studio. (787ed60)
Add multi-bank access and new MCP tools for interacting with multiple memory banks via MCP. (6b5f593)
Allow supplying custom entities when retaining memories via the retain endpoint. (dd59bc8)
Enhance the /reflect endpoint with max_tokens control and optional structured output responses. (d49e820)

Improvements

Improve local LLM support for reasoning-capable models and streamline Docker startup for local deployments. (eea0f27)
Support operation validator extensions and return proper HTTP errors when validation fails. (ce45d30)
Add configurable observation thresholds to control when observations are created/updated. (54e2df0)
Improve graph visualization to the control plane for exploring memory relationships. (1a62069)

Bug Fixes

Fix MCP server lifecycle handling so MCP lifespan is correctly tied to the FastAPI app lifespan. (6b78f7d)

0.1.15

Features

Add the ability to delete documents from the web UI. (f7ff32d)

Improvements

Improve the API health check endpoint and update the generated client APIs/types accordingly. (e06a612)

0.1.14

Bug Fixes

Fixes the embedded “get-skill” installer so installing skills works correctly. (0b352d1)

0.1.13

Improvements

Improve reliability by surfacing task handler failures so retries can occur when processing fails. (904ea4d)
Revamp the hindsight-embed component architecture, including a new daemon/client model and CLI updates for embedding workflows. (e6511e7)

Bug Fixes

Fix memory retention so timestamps are correctly taken into account. (234d426)

0.1.12

Features

Added an extensions system for plugging in new operations/skills (including built-in tenant support). (2a0c490)
Introduced the hindsight-embed tool and a native agentic skill for embedding/agent workflows. (da44a5e)

Improvements

Improved reliability when parsing LLM JSON by retrying on parse errors and adding clearer diagnostics. (a831a7b)

Bug Fixes

Fixed structured-output support for Ollama-based LLM providers. (32bca12)
Adjusted LLM validation to cap max completion tokens at 100 to prevent validation failures. (b94b5cf)

0.1.11

Bug Fixes

Fixed the standalone Docker image and control plane standalone build process so standalone deployments build correctly. (2948cb6)

0.1.10

This release contains internal maintenance and infrastructure changes only.

0.1.9

Features

Simplified local MCP installation and added a standalone UI option for easier setup. (1c6acc3)

Bug Fixes

Fixed the standalone Docker image so it builds and starts reliably. (b52eb90)
Improved Docker runtime reliability by adding required system utilities (procps). (ae80876)

0.1.8

Bug Fixes

Fix bank list responses when a bank has no name. (04f01ab)
Fix failures when retaining memories asynchronously. (63f5138)
Fix a race condition in the bank selector when switching banks. (e468a4e)

0.1.7

This release contains internal maintenance and infrastructure changes only.

0.1.6

Features

Added support for the Gemini 3 Pro and GPT-5.2 models. (bb1f9cb)
Added a local MCP server option for running/connecting to Hindsight via MCP without a separate remote service. (7dd6853)

Improvements

Updated the Postgres/pg0 dependency to a newer 0.11.x series for improved compatibility and stability. (47be07f)

0.1.5

Features

Added LiteLLM integration so Hindsight can capture and manage memories from LiteLLM-based LLM calls. (dfccbf2)
Added an optional graph-based retriever (MPFP) to improve recall by leveraging relationships between memories. (7445cef)

Improvements

Switched the embedded Postgres layer to pg0-embedded for a smoother local/standalone experience. (94c2b85)

Bug Fixes

Fixed repeated retries on 400 errors from the LLM, preventing unnecessary request loops and failures. (70983f5)
Fixed recall trace visualization in the control plane so search/recall debugging displays correctly. (922164e)
Fixed the CLI installer to make installation more reliable. (158a6aa)
Updated Next.js to patch security vulnerabilities (CVE-2025-55184, CVE-2025-55183). (f018cc5)

0.1.3

Improvements

Improved CLI and UI branding/polish, including new banner/logo assets and updated interface styling. (fa554b8)

0.1.2

Bug Fixes

Fixed the standalone Docker image so it builds/runs correctly. (1056a20)