Changelog

0.5.2 (April 24, 2026)

new features

• memory intelligence workbench — added query briefing, query comparison, and recent hotspot surfacing to the web dashboard so Engram can summarize what matters instead of only exposing raw tables and graph views
• 3 new MCP tools — focus_brief, compare_queries, and hotspots expose the same higher-level intelligence surface to agent clients over MCP

fixes

• postgres-safe drift dashboard — web drift check and drift fix now avoid SQLite-only PRAGMA busy_timeout and BEGIN IMMEDIATE calls when Engram is running on Postgres
• drift metadata compatibility — invalidated-memory drift checks no longer rely on SQLite json_extract(...), so the drift panel works correctly against Postgres jsonb
• better drift API errors — drift routes now return clean JSON errors for unexpected backend failures instead of surfacing broken frontend syntax errors
• postgres compatibility sweep — neural map aggregation, bridges, community summaries, quality metrics, and benchmark helpers now avoid SQLite-only query functions on Postgres-backed installs

0.5.1 (April 24, 2026)

fixes

• postgres migration hardening — engram migrate-postgres now tolerates stale SQLite references during real-world migrations instead of failing on orphaned entity_mentions, importance_history, or status_history rows
• migration verification — post-copy verification now compares against valid migratable rows, so installations with historical SQLite inconsistencies can still complete a clean verified move to postgres

0.5.0 (April 24, 2026)

new features

• postgres backend — engram can now use postgres as its primary storage backend for concurrent web + MCP deployments, while keeping sqlite as the default local-first option
• sqlite to postgres migration command — new engram migrate-postgres command copies an existing sqlite store into postgres, rebuilds full-text search rows, verifies key row counts, and can switch config.yaml automatically
• config-level backend selection — storage_backend and postgres_dsn are now first-class config fields and can also be driven by ENGRAM_STORAGE_BACKEND / ENGRAM_POSTGRES_DSN

docs

• added postgres migration guide
• updated install, config, CLI, and README docs for mixed sqlite/postgres deployments

0.4.1 (April 23, 2026)

fixes

• web dashboard stability — high-traffic dashboard routes now use short-lived store connections instead of the shared app store connection, which reduces SQLite contention under concurrent polling
• drift check + fix responsiveness — web drift operations now run off the async request thread, return fast 503 errors on lock contention, and avoid pinning the whole UI while a check or fix runs
• lighter dashboard boot — the web UI defers more work until it is needed and reduces background polling pressure during startup

docs

• release metadata updated for the 0.4.1 patch release

0.4.0 (April 23, 2026)

new features

• retrieval explainability — search now exposes query intent, expansion terms, exact-match boosting, cache-hit status, and candidate counts through recall_explain and /api/search/explain
• query expansion + exact match boosts — retrieval can expand common operator queries (auth, deploy, code, graph, memory) and reward exact phrase/token matches without overwhelming the dense pipeline
• search result caching — repeated hybrid searches are cached in-memory with automatic invalidation on writes, edits, forgets, layer changes, and bulk updates
• richer continuity tooling — session_checkpoint adds explicit stop-point checkpoints on top of automatic handoff refreshes, and resume_context now falls back to generating a handoff when no saved packet exists yet
• continuity APIs — the web server now exposes /api/session-handoffs and /api/session-handoffs/:session_id for dashboard and agent clients that want resumable session state without MCP

docs

• MCP tools reference now covers recall_explain and session_checkpoint
• REST API reference now documents the search explain endpoint and session handoff endpoints
• release metadata updated for the new feature release

0.3.2 (April 23, 2026)

new features

• structured session handoffs — active MCP sessions now maintain a resumable handoff packet with recent work, decisions, open loops, touched entities, and recall history
• 2 new MCP tools — session_handoff builds and optionally persists a structured handoff packet, resume_context loads the latest saved handoff for fast startup
• session continuity skill — added examples/skills/session-continuity/SKILL.md so agents have a concrete default pattern for startup, in-session writes, and stop-point handoffs

docs

• docs site now covers the session continuity flow on the home page, MCP tools reference, and Claude Code setup guide
• README and examples docs now point to the new continuity skill and handoff workflow

0.3.1 (April 14, 2026)

new features

• LLM API backends — anthropic and openai backends for fact extraction, memory enrichment, consolidation. configure with llm.backend + llm.api_key in config.yaml or env vars. pip install 'engram-memory-system[anthropic]' or [openai]
• memory types — fact, procedure, narrative column on every memory. indexed, filterable. auto-backfilled from existing metadata on migration
• retrieval profiles — recall accepts mode: facts_only (structured knowledge only), facts_plus_rules (+ procedures), full_context (everything). filters before cross-encoder reranking
• status tracking — lifecycle states: active, challenged, invalidated, merged, superseded. status_history audit table with timestamps and reasons. non-active memories excluded from retrieval
• 3 new MCP tools: update_status, recall_by_type, status_history (66 total)
• remember tool accepts memory_type param. remember_decision, remember_error, remember_project, remember_negative auto-set the correct type

fixes

• hnswlib knn_query crash when MCP passes float k (JSON numbers are floats)
• schema init on existing DBs — tolerates missing columns, runs migration before index creation
• config.yaml gitignored (contains API keys), config.example.yaml tracked

0.2.0 (April 11, 2026)

new features

• HNSW ANN index — approximate nearest neighbor via hnswlib, 100% recall@10, 0.09ms search, scales to 1M vectors
• 98.1% R@5 on LongMemEval — new SOTA, beating MemPalace (96.6%) and all others
• multi-backend embeddings — Voyage AI, OpenAI, Google Gemini alongside local MLX/sentence-transformers
• Voyage cloud reranker — rerank-2.5/2.5-lite alongside local cross-encoder
• SSE MCP transport — engram serve --mcp-sse for HTTP clients
• CLI: reembed — re-embed all memories after switching embedding model
• CLI: watch — poll a directory for new files and auto-ingest
• CLI: export/import — portable JSON backup with optional embeddings
• web auth — bearer token auth via web.auth_token config
• Docker — Dockerfile + docker-compose.yml
• 72 pytest tests across 6 modules
• 13 examples — setup guides + Python scripts for every major feature
• GitHub Actions CI — tests on push/PR, auto-publish to PyPI on release
• PyPI — pip install engram-memory-system
• docs site — MkDocs Material at engram-memory.dev

fixes

• ANN count tracks active ids (hnswlib mark_delete doesn't decrement)
• debug mode with rerank=False no longer crashes (unbound reranked variable)
• 9-factor importance (was incorrectly documented as 7-factor)
• 63 MCP tools (was incorrectly documented as 52)

0.1.0 (April 9, 2026)

initial release.

• 5-channel hybrid retrieval (dense + BM25 + graph BFS + Hopfield + RRF)
• memory layers (working, episodic, semantic, procedural, codebase)
• entity graph with co-occurrence relationships
• surprise-based importance scoring at write time
• retention regularization (L2/Huber/elastic)
• deep MLP reranker trained on access patterns
• dream cycle consolidation
• drift detection and auto-fix
• pattern extraction from sessions
• negative knowledge
• enriched embeddings (A-Mem)
• memory evolution
• intent-aware retrieval (MAGMA)
• trust-weighted decay
• 63 MCP tools
• web dashboard with neural map
• MLX GPU embedding backend