Changelog
All notable releases for AutoMem (the memory server) and MCP AutoMem (the MCP client) are documented here.
June 2026
Section titled “June 2026”0.16.0 · Recall Ranking Overhaul
Section titled “0.16.0 · Recall Ranking Overhaul”2026-06-26
A recall-ranking overhaul reworks how results are scored and ordered — the theme is correctness over knobs. Date-aware ranking prioritizes recent facts while keeping older context reachable, and latest-fact selection lets superseded memories give way to the current answer. The recency decay window and curve are now tunable, alongside an optional relative-recency re-rank, a relevance gate that requires topical evidence within the tag scope before applying query-independent scoring, and an opt-in tag-score cap that removes query-length bias. Recall gained state_mode=current|history to return only current memories or include superseded/invalidated ones for audit timelines, plus metadata sidecar search with metadata and updated_at surfaced in the detailed format. A single API call can now create or update many memory associations at once, a new admin endpoint exports memories for backup, and consolidation gained entity deduplication and identity synthesis with configurable cluster thresholds. /enrichment/status exposes the classification fallback-rate, and a new Recall Quality Lab harness tunes scoring against a clone of real data. See the Recall Tuning guide.
MCP 0.15.0 · Hermes, Guided Cloud & API-URL Rename
Section titled “MCP 0.15.0 · Hermes, Guided Cloud & API-URL Rename”2026-06-26
Two new install paths: AutoMem now installs into the Hermes terminal agent, and a guided cloud installer can stand up a hosted backend (InstaPods or Railway) and capture the endpoint and token for you. claude-code installs as an auto-updating Claude Code plugin that bundles the MCP server and hooks, replacing manual settings-level wiring. The endpoint variable AUTOMEM_ENDPOINT was renamed to AUTOMEM_API_URL (the old name still works as a fallback). The Stop hook now uses an LLM-judged storage nudge instead of mechanically capturing build/test/deploy output, store_memory can supersede an existing memory in one step, and recall responses are summary-first and token-budgeted to stay under the response-size cap. The installer’s permission scope was narrowed to the six mcp__memory__* entries. See What’s New.
April 2026
Section titled “April 2026”0.15.2 · Recall Quality & Tag Ops
Section titled “0.15.2 · Recall Quality & Tag Ops”2026-04-23
Keyword scoring for vector results with a softer adaptive floor — recall quality improves on short and keyword-heavy queries without drowning out semantic matches. Expansion candidates now bypass the tag filter so graph walks aren’t cut off at the seed set. New paginated GET /memory/by-tag with offset + has_more, plus a bulk DELETE /memory/by-tag for clearing stale scopes. LoCoMo judge runs hardened against flaky rate limits so benchmark sweeps complete cleanly.
MCP 0.14.0 · OpenClaw Plugin-First & Antigravity
Section titled “MCP 0.14.0 · OpenClaw Plugin-First & Antigravity”2026-04-23
Plugin-first memory parity for OpenClaw with a shared policy module across plugin, MCP, and skill modes — all three installation paths now surface the same typed tools. Google Antigravity setup templates added as a first-class platform. The standalone Claude Code plugin is deprecated in favor of npx @verygoodplugins/mcp-automem claude-code (see DEPRECATION.md). Windows users get a python-command.sh wrapper so hooks resolve the correct interpreter. Hooks hardened with stdin parsing, truncation guards, and dedup checks.
March 2026
Section titled “March 2026”0.15.1 · Deploy Freshness & Classifier Endpoints
Section titled “0.15.1 · Deploy Freshness & Classifier Endpoints”2026-03-25
OPENAI_BASE_URL now honored by the classification client, letting Azure and self-hosted OpenAI-compatible endpoints back the memory type classifier. Embedding provider dimension is validated at init time against the Qdrant collection — config drift fails fast instead of corrupting vectors. Consolidation cluster/creative embeddings now fetch from Qdrant directly. Railway deploy-freshness check guards against stale container rollouts.
0.15.0 · Relationship Engine & Benchmarks
Section titled “0.15.0 · Relationship Engine & Benchmarks”2026-03-25
Optimized relationship taxonomy for cleaner graph traversal. LoCoMo cat5 multi-hop judge for automated recall quality scoring. priority_ids now boosts relevance instead of hard-filtering. Embedding dimension validation catches provider mismatches at startup.
MCP 0.13.0 · OpenClaw Overhaul & Cursor Rules
Section titled “MCP 0.13.0 · OpenClaw Overhaul & Cursor Rules”2026-03-25
Complete OpenClaw plugin rewrite — native MCP integration and new skill setup. Split global vs. project rules for Cursor. Hook stdin parsing with truncation guards and dedup checks.
0.14.0 · Docker & Hardened Deployments
Section titled “0.14.0 · Docker & Hardened Deployments”2026-03-08
Official Docker build workflow. QDRANT_HOST + QDRANT_PORT for flexible Qdrant config. Stateless MCP bridge transport for resilient connections. Consolidation overhaul: reduced decay rate, importance floor, archived memory filtering. Voyage AI set as recommended embedding default.
0.13.0 · Adaptive Recall & Graph Viewer
Section titled “0.13.0 · Adaptive Recall & Graph Viewer”2026-03-02
min_score threshold with adaptive floor filtering — low-relevance results get cut automatically. LoCoMo benchmark harness for rapid experiment iteration. Graph viewer externalized as standalone module.
February 2026
Section titled “February 2026”0.12.0 · Recall Quality Lab & JIT Enrichment
Section titled “0.12.0 · Recall Quality Lab & JIT Enrichment”2026-02-20
Just-in-time enrichment on recall for higher-quality results. Recall Quality Lab for data-driven recall optimization. LongMemEval benchmark harness (ICLR 2025) for measuring memory quality. relevance_score now synced to Qdrant payload for accurate vector search.
0.11.0 · Embedding Providers & Memory Lookup
Section titled “0.11.0 · Embedding Providers & Memory Lookup”2026-02-16
Voyage AI embedding provider for high-accuracy recall. OpenAI-compatible providers via OPENAI_BASE_URL (LM Studio, Ollama, xAI, etc.). GET /memory/<id> endpoint for direct memory lookup. Hardened time parser for numeric timestamps.
0.10.1 · Stability Fixes
Section titled “0.10.1 · Stability Fixes”2026-02-07
Temporal query bound to 7-day window with timeout to prevent runaway queries. Skip temperature param for o-series and gpt-5 models.
0.10.0 · Visualization & Observability
Section titled “0.10.0 · Visualization & Observability”2026-02-06
Interactive graph visualization API. automem-watch real-time observability with SSE streaming and TUI. Ollama embedding provider for fully local deployments. Memory content size governance with auto-summarization. Streamable HTTP transport for MCP connections.
MCP 0.12.0 · OpenClaw Integration
Section titled “MCP 0.12.0 · OpenClaw Integration”2026-02-06
Native OpenClaw integration replacing mcporter dependency chain. Direct HTTP API calls to AutoMem with new SKILL.md template. Simplified setup and troubleshooting documentation.
MCP 0.11.0 · Claude Code Plugin
Section titled “MCP 0.11.0 · Claude Code Plugin”2026-02-03
Claude Code plugin with automatic memory capture hooks. Session start recall, session end summary. Cursor hooks integration for automatic capture. Integration tests and CLI smoke tests.
January 2026
Section titled “January 2026”MCP 0.10.0 · MCP Registry & SessionStart Hook
Section titled “MCP 0.10.0 · MCP Registry & SessionStart Hook”2026-01-06
Listed on the official MCP Registry. SessionStart hook for automatic memory recall at conversation start. Memory content size guidelines and warnings surfaced in the client.
December 2025
Section titled “December 2025”0.9.3 · Expansion Filtering
Section titled “0.9.3 · Expansion Filtering”2025-12-10
expand_min_strength and expand_min_importance parameters for fine-grained graph traversal control. Vector dimension autodetect for embedding providers. CI workflow for automated testing on PRs.
MCP 0.8.1 · Spec Compliance
Section titled “MCP 0.8.1 · Spec Compliance”2025-12-04
All MCP tools now return structuredContent with outputSchema, aligning recall outputs (results, dedup_removed) to spec.
MCP 0.8.0 · Advanced Recall Tools
Section titled “MCP 0.8.0 · Advanced Recall Tools”2025-12-02
Exposed expand_entities/relations, auto_decompose, context-aware boosts, and full tool metadata in the MCP client. Simplified Claude Code integration.
0.9.1 · Entity Expansion + Benchmarking
Section titled “0.9.1 · Entity Expansion + Benchmarking”2025-12-02
Added expand_entities/entity_expansion params for graph hops and recorded an early LoCoMo-10 benchmark milestone. Current benchmark claims are maintained on the generated Benchmarks page.
November 2025
Section titled “November 2025”0.9.0 · Retrieval Engine Upgrade
Section titled “0.9.0 · Retrieval Engine Upgrade”2025-11-20
Multi-hop bridge discovery, temporal alignment scoring, and weighted hybrid scoring (vector, keyword, relation, temporal, importance) boost recall precision.
0.8.0 · API Modularization & Security
Section titled “0.8.0 · API Modularization & Security”2025-11-08
Refactored monolithic app into modular blueprints, hardened memory IDs to server-generated UUIDs, and fixed Railway secrets/volumes for reliable deploys.