Changelog

0.8.1 (September 20, 2026)

• suppress only the known BGE reranker and default embedding reports about regenerated position_ids buffers; preserve other loading diagnostics and remove blanket Transformers warning suppression
• refresh engram demo around a fictional project, current retrieval explanations and setup commands, with unattended execution and isolated demo storage
• fix demo compatibility with Python 3.11 and ensure temporary data and web processes are cleaned up on exit
• carry the web page's access token into same-origin API requests so authenticated workspaces load and search correctly

0.8.0 (September 20, 2026)

• add guided engram init with local model presets, private new-store creation, environment-override reporting and portable MCP connection settings
• add engram doctor for configuration, storage and model readiness; --full verifies real model inference, a local MCP handshake and isolated save/retrieve behavior
• add serve --mcp --no-warmup so endpoint diagnostics can inspect the protocol without loading models
• retry a focused excerpt for confidence-rejected long memories even when another candidate blocks the early retry; preserve the existing confidence cutoff and one-excerpt-per-document limit
• omit repeated unrelated boilerplate next to a complete matching fact while retaining unique context and recognized qualifications
• add reproducible synthetic production-retrieval checks, separate from the historical LongMemEval result

0.7.0 (September 20, 2026)

• validate configuration before startup, reject unknown fields and malformed values, and support an environment override for every setting
• add engram config check, show and schema, with redacted values and provenance; inspection does not initialize storage or load models
• share effective settings through MCP config_show, native config_show, and the authenticated web configuration endpoint
• show why retrieval candidates were returned, rejected by confidence, filtered or excluded by the result budget; include model scores and focused-excerpt traces
• make explanation mode bypass result caching and leave access history, dormant evaluations and MCP session handoffs unchanged
• expose the shared report through CLI --explain (--debug remains an alias), native search_explain, MCP recall_explain, and the web workspace
• use the configured result limit when ordinary search or its explanation omits one, while retaining interface input limits
• keep machine-readable CLI output clean when model libraries print diagnostics
• honor a zero cache capacity as disabled, synchronize resolved Hugging Face credentials, and recheck memory eligibility after reranking
• redesign the logo, public website and documentation around a dark copper identity, with responsive layouts, keyboard navigation and reduced-motion support
• centralize interface version reporting and verify built package contents before publication
• independently confirm 470/470 session recall-any@5 in a fresh full LongMemEval development run without saved-score reuse; publish text-free question rankings and provenance, with abstention, confidence and native API limitations
• correct the comparison table: label external answer-accuracy and retrieval metrics separately, link primary reports and remove unsupported misses-out-of-470 claims
• use BAAI/bge-reranker-base as the default local reranker; keep MiniLM available through config
• return raw local reranker logits and apply a sigmoid once; preserve hosted normalized scores
• retry at most one focused source excerpt per eligible long document when every full-document local sigmoid score is below rerank_passage_floor (default 0.001); keep the 0.6 production confidence gate independent and retain max full/excerpt aggregation
• match conservative regular English singular/plural forms during excerpt selection, counting each original query term once per sentence
• retain full/excerpt raw scores and source offsets; keep the semantic query unchanged and exclude resolved relative-time text only from lexical selection
• remove extra lexical/prior bonuses and random noise from cross-encoder results; honor the optional bounded rerank_fusion_alpha blend
• keep the best confidence-eligible hybrid candidate in requests for at least two results, without changing scores or replacing the rerank winner; expose preserve_prior_candidate to disable coverage
• separate cached rerank results when the fusion weight, confidence gate, coverage setting, passage fallback flag or activation floor changes
• preserve query spelling instead of applying a single-word substitution
• share date parsing and relative windows with the benchmark, and apply its temporal boost once per session
• correct NDCG rank discounts and include unretrieved answers in the ideal ranking
• record benchmark dataset/source hashes, model names, result window, passage fallback flag/floor and settings; support --no-passage-fallback and --passage-floor, and reject incompatible or duplicate resume records
• preserve configured fusion and passage activation values unless the corresponding CLI option overrides them; document the activation floor as a benchmark development choice, not a calibrated probability or held-out accuracy claim
• refuse existing output paths for fresh benchmark runs and claim the result file before writing new provenance

the fresh benchmark above measured the retrieval candidate published in c3b8cea, before the configuration and explanation additions. the evidence keeps its original source fingerprints; the complete 0.7.0 package was not given a second full LongMemEval run.

0.6.2 (September 19, 2026)

• add hf_token config and env fallback so model downloads authenticate without manual huggingface-cli logins
• resolve relative dates ("last Saturday", "N days ago") and keep the temporal boost after cross-encoder reranking
• switch license to Engram Public Use License 1.0

0.6.1 (September 18, 2026)

• widen cross-encoder candidate pool from 20 to 35 so deep semantic hits don't get buried by lexical BM25
• use regex tokenization in BM25 so trailing punctuation doesn't kill keyword matches

0.6.0 (September 18, 2026)

new features

• dormant recall in shadow mode — tracks queries with low confidence or empty retrieval results in an associative shadow buffer; when new memories are ingested or consolidated, the system automatically checks for newly satisfied associations and reactivates dormant queries (engram dormant CLI, config controls, and background evaluation)
• native memory service layer — introduced EngramService (engram/service.py), providing a high-level programmatic interface that unifies memory CRUD, evidence tracking, project context assembly, and hybrid retrieval
• evidence graph & provenance tracking — added first-class evidence extraction and verification (engram/evidence.py) to trace ground-truth source references, causal antecedents, and factual reliability
• project context provider — added session-aware context distillation (engram/project_context.py) for assembling working memory state, open decisions, and active task parameters for coding agents
• codex adapter — added native adapter (engram/adapters/codex.py) bridging Codex tool calls, memory handoffs, and skill execution directly into Engram
• neural graph & connections overhaul — complete overhaul of the web dashboard's Connections workspace into an interactive neural exploration canvas:
• 2D camera with cursor-anchored zoom (0.25x–4.0x), smooth click-drag panning, and reset controls
• dual layout modes: concentric layer rings (Working → Semantic → Episodic → Procedural) with glowing orbits and force-directed spring cluster physics
• constellation mode: click-to-spotlight 1-hop and 2-hop subgraphs with camera targeting and unrelated node dimming
• in-graph inspector drawer showing direct entity connections with jump links, linked memory excerpts, and inline type/alias management
• live entity search with auto-suggest dropdown and keyboard navigation (/, Esc, Enter)
• real-time synaptic pulse cascades with shockwave bursts and edge particle flows (⚡ pulse button, F hotkey, and live access log polling)
• dual sizing metric combining memory mention volume and relationship degree
• zigcho-infra design language — full UI reskin of the dashboard to the refined slate-black palette (#09090a base, #111113 surface, #292427 borders, #ed83b6 signature pink, #8fd4b1 mint, #7bbcff blue)
• unbounded exports — /api/export now supports exporting all memories without forced pagination caps

fixes

• cross-database SQL compatibility — replaced postgres-specific cast syntax with standard CAST(e.aliases AS TEXT) in entity search so MCP entity discovery runs safely across SQLite and Postgres
• defensive consolidation — access log and event pruning now defensively handles differing driver rowcount behaviors across database backends
• web script parsing — eliminated duplicate state declaration in the web dashboard preventing browser reference errors
• standard package metadata — pinned build-system hatchling to ensure strict compliance with packaging metadata standards across build environments

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
• historically reported 98.1% session recall-any@5 on a 470-question LongMemEval subset; see measurement scope and limitations
• 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