Skip to content
Agents tracked: 501 Releases (7d): 777 Agent status: no reported issuesSince last limit reset: Claude 10d · Codex 8d For AI agents: Markdown Updated Nov 21, 2026

Hindsight changelog: what's new each month

Every stable Hindsight release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 12 months covered; the current month updates daily.

November 2026 · so far

4 releases: 1.2.0 → 1.2.3

November brought stability improvements and new capabilities to Hindsight, with multimodal document banks arriving behind a feature flag and several important fixes to core operations. The month focused on robustness, particularly around file handling and API reliability.

Highlights

  • Multimodal document banks now supported, allowing richer knowledge storage.
  • New reflect_budget feature lets you balance reflection depth against recall performance independently.
  • File ingestion now handles malformed, unusual and large files without failing or consuming excessive memory.
  • Cron jobs and kanban workers now operate more efficiently with skip-auto-recall and read-only modes.
  • API responses are now more accurate, with proper HTTP status codes for LLM provider content refusals.

New

  • Multimodal document banks for knowledge storage (feature-flagged).
  • reflect_budget parameter to tune reflection behaviour separately from recall.
  • Cron jobs can skip auto-recall to reduce unnecessary processing.
  • Kanban workers can operate in read-only mode without writing changes.

Improved

  • File ingestion is now more robust, handling odd, broken and large files without failure.
  • Session memory append operations no longer rewrite all memories unnecessarily.
  • Reranker now correctly honours local FP16 setting on CPU.
  • Admin reembed command distinguishes leftover columns by model and dimension.
  • Survey on coding-agents now starts from every host and preserves terminal environment settings.
  • Client disconnect no longer incorrectly counts as failed operations in metrics.

Fixed

  • UTF-8 byte-order mark in markitdown text files is now stripped correctly.
  • Embeddings API no longer fails when provider omits prompt_tokens.
  • Reflect operations handle non-list id and memory_ids fields gracefully.
  • Windows redirected stdout no longer stops server startup.
  • API now returns 422 instead of 500 when LLM provider refuses content.
  • Double gzipping of /api/graph on Next 16.4.6 has been removed.

October 2026

3 releases: 0.10.3 → 1.1.0

October brought Hindsight to version 1.0.0 and beyond, with improvements to agent memory handling, new extension capabilities, and fixes to stability and performance across the platform.

Action needed
  • Agent SDK package and agent_knowledge_* tools have been removed.

Highlights

  • LLM request extensions now let you choose the model and credentials for each call server-side.
  • Memory is now treated as a past record that teams can add their own guidance to.
  • Webhook event delivery now includes event IDs for better tracking and reliability.
  • Fixed the deepen lock mechanism to use heartbeats instead of a 30-minute timeout, improving reliability.
  • Reached version 1.0.0 milestone with documentation snapshot and Memlake launch.

New

  • TraeCode (TRAE CN) harness support for coding agents.
  • LLM request extension for server-side control of model and credential selection.
  • Webhook event delivery IDs for tracking event delivery.
  • Control-plane requests now identify themselves with their own User-Agent.
  • Hermes daemon now runs out-of-process instead of being imported inline.

Improved

  • Memory in coding agents is now treated as a past record with team-added guidance.
  • Export poll loop no longer overruns its timeout on Windows.
  • Mental-model refresh reflect calls now run under their own concurrency cap.
  • Bank deletion is now a no-op when the bank does not exist instead of returning an error.
  • Non-conversation Codex content is filtered out in coding agents.
  • Observations only rebuild when tag changes actually move their scopes.

Fixed

  • Deepen lock now uses heartbeats instead of a 30-minute TTL for better stability.
  • Tag-scoped callers are now refused on operations answered for the whole bank.
  • Documentation homepage link preview now displays correctly.
  • Wide-screen documentation hero now reaches the window edges.
  • No-op '+field' button is hidden at the entity-label root level.
  • Embedded Hermes daemon stability improved by running out-of-process.

September 2026

3 releases: 0.10.0 → 0.10.2

September brought performance improvements to Hindsight's core memory and search systems, along with better API attachment handling and more reliable coding agent behaviour. Documentation was overhauled with tested code examples across the board.

Highlights

  • Replaced tiktoken with quicktok for faster tokenization, defaulting to o200k encoding.
  • Fixed batch processing for document chunk embeddings to reduce API requests.
  • All code examples in documentation are now tested and verified.
  • Improved attachment handling for store-owned banks with proper reclamation guards.
  • Coding agents now use concise extraction by default with configurable caching.

New

  • Attachment filenames are now supported for store-owned banks.
  • Memory defense redacts Hindsight Cloud API keys from logs.
  • Coding agents can now append to cache during concurrent operations.

Improved

  • Document body accumulation no longer rebuilds per sub-batch, improving memory efficiency.
  • Oversized items are rebuilt from original spans instead of sub-batches.
  • Full-text search results are now scored per row for better ranking.
  • Store lock no longer fails when the store cannot be written during OAuth.
  • Mental models now refuse off-schema delta replies and re-request from the model.
  • Concise prompt examples in retain mode are now guarded.
  • Bank list pages now order by the store instead of tenant ranking.
  • Codex agent turns are now correctly extracted from UserMessage events.

Fixed

  • Prevent a single stray file from aborting Claude import in coding agents.
  • Fixed benchmark imports of removed engine symbols.
  • Attachments are never reclaimed for store-owned banks.
  • Docs skill's OpenAPI specification is now synced to current version.
  • Improved documentation now shows tested live snippets and fixed four issues they revealed.
  • Versioned docs for version 0.6 have been removed.

August 2026

3 releases: 0.9.0 → 0.9.2

August brought pluggable memory storage, client-managed knowledge pages, and numerous stability improvements across the platform. The month focused on enhancing flexibility and fixing issues in concurrent operations, database handling, and search backends.

Highlights

  • Pluggable memory storage backend lets you choose where Hindsight stores agent memories.
  • Client-managed knowledge pages via control-plane UI and CLI give you direct control over knowledge bases.
  • Hindsight-hermes standalone plugin is now deprecated in favour of integrated functionality.
  • Fixed critical deadlocking issues in concurrent operations and vector-index management.
  • Improved search reliability across different text-search backends.

New

  • Pluggable memories storage backend for flexible memory persistence.
  • Client-managed knowledge pages with control-plane UI.
  • Hindsight filesystem CLI for knowledge base management.

Improved

  • Knowledge-base search now works reliably on non-native text-search backends.
  • Coding agents now properly seed the harness instead of using opencode defaults.
  • Graph maintenance is now serialised per bank at claim time for better consistency.
  • Entity postings and witness coverage handling improved within write-groups.
  • BM25 long query search keeps the most selective terms for better relevance.
  • Database connection pooling now preserves application names across reuse.
  • LLM integration forwards extra body parameters to Codex requests.
  • Fallback worktree handling for coding agents improved from filesystem-root.

Fixed

  • Concurrent bank deletes no longer deadlock on vector-index DDL operations.
  • PGroonga Compose Docker image repair.
  • Knowledge-base search 500 errors on non-native text-search backends.
  • Relink victims now enqueued before delta chunk deletion in retention.
  • Test fixtures no longer delete other xdist workers' operations in schema claims.

July 2026

3 releases: 0.8.4 → 0.8.6

July brought improved integrations with AI development tools, fixes to memory filtering and display, and enhancements to handle larger datasets and prevent data loss. The month focused on stability and expanding platform support.

Highlights

  • Added Aider integration with session-bracketing memory wrapper for better development workflow continuity.
  • Fixed mental model display to show all results instead of capping at 100, improving discoverability.
  • Implemented memory filtering by ingest age to help manage large memory collections.
  • Fixed cosine similarity calculations for non-unit embeddings to improve memory relevance ranking.
  • Added persistent memory support for Devin Desktop and improved control-plane UI stability.

New

  • Aider integration for session-bracketing memory wrapper
  • Filter memory units by ingest age (createdbefore parameter)
  • Custom headers support in TypeScript client
  • Strict schema enforcement via forced tool call and repair in Codex

Improved

  • Mental models view now displays all results without capping at 100
  • Memory filters now remain visible when displaying empty results
  • JSON viewer now correctly displays Unicode output
  • Chunk text serialization made idempotent to handle raised structural chunk sizes
  • Cosine similarity calculations corrected for non-unit embeddings
  • Documentation updated to reflect current fact types and integrations

Fixed

  • Fixed mental model display cap that was hiding results beyond 100
  • Fixed JSON viewer Unicode output display in control-plane
  • Fixed cosine similarity calculations for non-unit embeddings
  • Fixed chunk text handling to prevent failures when structural chunk size increases
  • Fixed memory filter visibility on empty search results
  • Resolved 62 high-severity dependency alerts

June 2026

5 releases: 0.7.2 → 0.8.3

June brought significant expansion of Hindsight's integration ecosystem and improved core functionality across memory retention, retrieval, and tracing. The month saw new support for multiple AI code editors and embedding models, alongside critical stability and performance fixes.

Highlights

  • Hindsight now integrates with Gemini Spark, Cursor, and Cline as installable plugins for persistent memory in these editors.
  • Added per-bank LLM request tracing via OpenTelemetry for better observability of memory operations.
  • Improved memory recall with per-source candidate caps and better freshness checking to prevent stale results.
  • MiniMax default model upgraded to M3 for better performance.
  • Gemini Embedding 2 family support added for per-input embedding capabilities.

New

  • Hindsight integration for Gemini Spark via MCP protocol.
  • Cursor editor integration with memory plugin and CLI tool.
  • Cline integration with persistent memory via lifecycle hooks (no MCP required).
  • Per-bank LLM request tracing using OpenTelemetry GenAI recorder.
  • Support for Gemini Embedding 2 family with per-input embedding.
  • Roo-code and Cursor CLI available as pip-installable packages.

Improved

  • MiniMax default model upgraded to M3.
  • Memory recall enhanced with per-source candidate caps on VectorChord BM25.
  • Pre-extraction freshness recheck added with concurrent write serialization for data integrity.
  • Search recency scoring now uses effective-time fallback (mentionedAt, occurredEnd) when primary timestamps unavailable.
  • Isolate Claude Code provider subprocess from user plugins to prevent interference.
  • Reduce false negatives in metrics by not counting client disconnections as failed recall/reflect operations.

Fixed

  • Memory recall now properly gates VectorChord BM25 results.
  • Fixed concurrent same-document writers to prevent data corruption.
  • Corrected recall folding in opencode to integrate into first system section instead of creating a new one.
  • Explicit daemon host and port configuration now properly honored.
  • Tokenizers dependency capped at 0.23.0 for local-ML extras to ensure compatibility.
  • Migration routines now install correctly on target schema public.

May 2026

5 releases: 0.6.0 → 0.7.1

May brought major infrastructure improvements and new AI backend integrations to Hindsight. The month saw the addition of Oracle 23ai support, AWS Bedrock AgentCore integration, and expanded language model provider options.

Action needed
  • Self-driving-agents CLI moved to separate repository.
  • Drop per-package CHANGELOGs in favor of single integration table.
  • OpenClaw drops redundant beforeagentstart configuration.

Highlights

  • Oracle 23ai database backend now supported alongside existing databases.
  • AWS Bedrock AgentCore Runtime integration for running agents on managed infrastructure.
  • TypeScript client now generates concrete types instead of generic Promise<any> for better type safety.
  • OpenCode-Go LLM provider added as a new language model option.
  • Memory efficiency improvements reduce pressure on retained content during agent operations.

New

  • Oracle 23ai database backend support.
  • Hindsight-AgentCore integration for Bedrock AgentCore Runtime.
  • OpenCode-Go language model provider.
  • Debug performance timing feature in OpenClaw.
  • Memory stats enrichment in control plane bank dropdown.
  • Category filter for blog landing page.
  • Claude-code create-agent skill now understands SDA layout.
  • GraphMaintenance operations type in control plane filter.

Improved

  • Memory efficiency by clearing content references after use in retain operations.
  • TypeScript client replaced generic Promise<any> with concrete generated types.
  • Claude-code getpage now handles detail=content and tool-result spillover.
  • OpenClaw mission semantics and retainQueue config whitelist refined.
  • Control plane Chinese translation improved for naturalness.
  • PGroonga added to BM25 backends table documentation.
  • Strands client lifecycle management improved to prevent leaks.
  • API now strips Gemma thought tags automatically.

Fixed

  • Broken link in CLI ARM64 guide.
  • Oracle query semantics restored and migration chain cleaned up.
  • OpenClaw silent dispatch skipping on synthetic-main and static-banking setups.
  • Paperclip integration aligned with actual event payloads.
  • Claude-code MCP parameter renamed from maxresults to maxtokens.
  • Markdown tables formatting in documentation.

April 2026

7 releases: 0.5.0 → 0.5.6

April brought significant improvements to Hindsight's memory and retrieval capabilities, with new support for multiple AI platforms and better performance handling. The release cycle focused on stability, configuration flexibility, and enhanced integration with external systems.

Action needed
  • Removed BFS and MPFP graph retrieval strategies.

Highlights

  • Added Constellation view with Pretext canvas rendering for better memory visualization.
  • Introduced session-scoped document IDs and per-message timestamps for more precise memory tracking.
  • New support for Anthropic-shaped JSON retention format with tool use and tool result blocks.
  • Added named connection profiles and per-operation slot reservations for improved configuration and performance.
  • Replaced Paperclip library with dedicated plugin for better modularity.

New

  • Constellation view in control plane with Pretext canvas rendering.
  • Session pattern filtering for ignore and stateless sessions in openclaw.
  • Config-aware history backfill CLI tool.
  • Named connection profiles (-p/--profile) for CLI.
  • Per-operation slot reservations for worker resource management.
  • System performance test runner.
  • Decommission-workers and worker-status CLI commands for admin operations.
  • Configurable Postgres statement timeout on pool connections.

Improved

  • 3-phase retain pipeline now fixes deadlocks, caps temporal links, and enables query-time entity expansion.
  • Observation quality improved with structured processing rules in consolidation.
  • Resilient startup and richer retain metadata in openclaw.
  • Session identity now stabilized and operational turns skipped appropriately.
  • Mental model max tokens properly forwarded to refresh in reflect.
  • Control plane bank ID encoding end-to-end in URLs.
  • Added tenant field and configurable allowlist to JSON logs.
  • Prevented child tasks from blocking parent execution in worker.

Fixed

  • Fixed timeline group sort to use numeric Date comparison instead of localeCompare.
  • Vector extension config now respected in per-bank index migration.
  • Addressed critical and high severity security vulnerabilities in dependencies.
  • Fixed deadlocks and stability issues in session identity handling.
  • Fixed agent hook registration on every plugin entry invocation.
  • Corrected session.messages response shape in opencode plugin.

March 2026

8 releases: 0.4.15 → 0.4.22

March brought significant performance improvements, expanded integration support, and critical security fixes to Hindsight. The release focused on enhancing memory management, adding new reranker options, and strengthening compatibility with popular agent frameworks.

Action needed
  • Fact type 'assistant' renamed to 'experience' across extraction pipeline.
  • Hardcoded default models removed from integrations.
  • LiteLLM 1.82.8 excluded due to supply chain compromise.

Highlights

  • New webhook system with retain.completed events for better integration with external systems.
  • Performance boost through HNSW indexing replacing window-function retrieval for faster memory recall.
  • Added support for multiple new rerankers including Jina-MLX for Apple Silicon devices.
  • Introduced configurable token limits for source facts during memory consolidation and recall.
  • Expanded integrations with LangGraph, LangChain, and Strands Agents SDK.

New

  • Webhook system with retain.completed event, UI, and documentation.
  • Observation scopes field to control memory observation granularity.
  • Source facts token limits for consolidation and recall operations.
  • Jina-MLX reranker provider for Apple Silicon support.
  • Configurable base URL for ZeroEntropy reranker.
  • Strategy parameter for MCP retain tool.
  • Independent versioning for integrations.
  • Linux ARM64 CLI binary release.

Improved

  • OpenClaw now retains last n+2 turns every n turns (default n=10) for better context management.
  • Bank dropdown now refreshes when opened and truncates long bank names in selector.
  • Recall max query tokens made configurable via environment variable.
  • MiniMax default model upgraded from M2.5 to M2.7.
  • Improved type hints and documentation in client wrapper.
  • 4-tab code parity across all documentation examples.
  • Better support for Gemini 3.1 Flash Lite with thought signature in tool calls.
  • Added scrolling integrations banner to all documentation pages.

Fixed

  • JSON serialization and logging exception propagation in Claude code LLM.
  • ZeroEntropy rerank URL missing /v1 prefix and MCP retain async processing parameter.
  • In-flight async operations now cancelled when bank is deleted.
  • Entity ID null constraint for non-ASCII entity names.
  • Hermes agent async client methods to prevent event loop deadlock.
  • MCP middleware now injects Accept header to prevent 406 errors.

February 2026

7 releases: 0.4.8 → 0.4.14

February brought expanded LLM support, improved async operations, and new integrations for Hindsight. The month focused on reliability fixes, better context propagation, and enhanced retrieval capabilities.

Highlights

  • Added support for Codex and Claude Code as LLM providers, plus dynamic llmProvider/llmModel configuration options.
  • Introduced CrewAI integration for persistent crew memory and new Go client SDK with automatic code generation.
  • Launched OpenClaw external Hindsight API support with dynamic per-channel memory banks and autoRecall toggle.
  • Added OpenTelemetry traceability and new docs skill for improved observability and knowledge access.
  • Support for Batch API for retain operations via OpenAI and Groq for more efficient processing.

New

  • Pre-operation validation hooks for mental model create and refresh operations.
  • Markdown support in reflect and mental models.
  • Chunks-only recall mode with maxtokens=0 parameter.
  • Source facts included in observation recall results.
  • ZeroEntropy reranker provider support.
  • Go client SDK with automatic code generation.
  • OpenTelemetry traceability for monitoring.
  • Docs skill for documentation access.

Improved

  • Request context now propagates through async task payloads for better traceability.
  • Shell argument escaping in OpenClaw improved for safer command execution.
  • Claude Code and Codex support improved for /reflect operations.
  • Async batch retain now handles large payloads more efficiently.
  • Temporal ordering offset reduced from 10 seconds to 10 milliseconds for faster operations.
  • MCP usage metering now correctly propagates tenant ID through context variables.
  • Helm charts now support per-component affinity and Pod Disruption Budgets.
  • Dynamic per-channel memory banks added to OpenClaw for flexible memory organization.

Fixed

  • Custom PostgreSQL schema handling is now more reliable.
  • Tagged directives are now correctly applied to tagged mental models.
  • Operation validator extension is properly loaded in worker processes.
  • Documents with zero extracted facts are now tracked correctly.
  • Client timeout settings are now properly respected.
  • Reranker no longer crashes on provider errors.

January 2026

11 releases: 0.2.0 → 0.4.7

January brought significant expansion to Hindsight's capabilities, adding support for multiple LLM providers, embedding customization, and new features for team collaboration and memory management. The month saw steady improvements across configuration, reliability, and multi-tenant support.

Highlights

  • Added support for Anthropic Claude, LM Studio, Groq, and Vertex AI as LLM providers, plus configurable OpenAI embeddings with adjustable dimensions.
  • Introduced mental models as a new capability for organizing and structuring agent memory.
  • Added cloud mode to skill installer to enable team memory sharing across users.
  • Implemented backup and restore functionality for preserving agent state.
  • Launched moltbot integration and real-time performance logging for consolidation operations.

New

  • Support for Anthropic Claude and LM Studio provider support.
  • Configurable observation thresholds for fine-tuning agent behavior.
  • Mental models feature for structured memory organization.
  • Cloud mode for skill installer enabling team memory sharing.
  • Backup and restore functionality.
  • Structured JSON logging support.
  • moltbot integration.
  • Vertex AI as LLM provider with native genai SDK.

Improved

  • Enhanced /reflect endpoint with max tokens and structured output options.
  • Improved causal links detection in memory consolidation.
  • Operation validator extension now integrated with proper HTTP error handling.
  • Configurable embedding dimensions for OpenAI embeddings.
  • Made max completion tokens retention configurable.
  • Added real-time timing breakdown logging for consolidation operations.
  • Enhanced LLM retry configuration options.
  • Improved multi-tenant schema handling and database setup at startup.

Fixed

  • Fixed Groq LLM support with free tier.
  • Fixed Python SDK not sending Authorization header.
  • Fixed MCP lifespan management with FastAPI app lifecycle.
  • Fixed hindsight-embed crashes on macOS.
  • Fixed /version endpoint returning wrong version number.
  • Fixed retain operations with timestamps failing asynchronously.

December 2025

19 releases: 0.0.16 → 0.1.16

December brought Hindsight through its 0.1.x phase with rapid iteration and foundational feature additions. The month focused on expanding model support, improving reliability, and introducing extensibility options.

Highlights

  • Added support for Gemini 3 Pro and GPT 5.2 models
  • Introduced optional graph retriever with MPFP capability
  • Launched local MCP server support for enhanced integration
  • Added extensions framework and native agentic skills
  • Improved LLM JSON parsing with retry logic and detailed error handling

New

  • Optional graph retriever MPFP capability
  • Local MCP server support
  • Support for Gemini 3 Pro and GPT 5.2 models
  • Hindsight-embed module with native agentic skill
  • Extensions framework with documentation
  • Document deletion functionality in the UI
  • Hindsight-liteLLM implementation for LLM integration
  • Health endpoint improvements

Improved

  • Upgraded Next.js to 16.0.10 to patch security vulnerabilities
  • Enhanced LLM JSON parsing with retry logic and detailed logging
  • Improved exception propagation from task handlers to enable retry logic
  • Switched to pg0-embedded for database handling
  • Refactored hindsight-embed architecture
  • Enhanced documentation and contributing guidelines
  • Improved Ollama structured support
  • Better test coverage and CI compatibility checks for Python 3.11, 3.12, and 3.13

Fixed

  • Bank selector race condition when switching between banks
  • Retain async failures
  • Timestamp not being considered in retain functionality
  • DOM.Iterable type errors for URLSearchParams
  • Ollama structured output support
  • Embed get-skill installer issue

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Hindsight: pricing, features and alternatives · All changelogs