MIND-Mem

MIND-Mem is a memory layer for AI agents that stores decisions, tasks, entities, and incidents as plain Markdown files on disk and exposes them through the Model Context Protocol. Recall is defined as a function of three inputs — corpus, configuration, and a scoring_instant UTC date — so the same three return the same ranked results on any host and any day; omitting the date resolves it to today in UTC and binds that value into the recall attestation, so an attested run replays by passing the date back. Ranking scores are ordinary floating point; the byte-identical guarantee applies to a Q16.16 fixed-point audit chain embedded in every applied decision.

Writes are governed rather than direct. Nothing reaches the source of truth without propose → review → approveapply, and governance modes graduate from detectonly to propose to enforce. Each apply is logged with a timestamp, a receipt, and a diff, taken against a snapshot with automatic rollback on validation failure, and a batch review command shows pending proposals with their pre-apply diff, provenance, and chain status.

Beyond storing and retrieving, an integrity layer looks for memory that has gone wrong: contradictions between decisions, drift from informal choices never formalized, dead decisions nothing references, and orphan task references, checked against 74 or more structural rules. Detected conflicts run through a graduated resolution pipeline — timestamp priority, confidence priority, scope specificity, then manual fallback — which generates supersede proposals and never applies without review.

Retrieval combines thread-parallel BM25F field-weighted scoring with optional vector search, fused by Reciprocal Rank Fusion, and adds RM3 query expansion, a nine-type intent router that maps each query type to its own retrieval parameters, two-hop graph neighbour boosting, deterministic four-signal reranking, and an optional cross-encoder. A pre-model confidence gate scored from entity overlap, BM25 score, speaker coverage, evidence density, and negation asymmetry forces abstention below threshold rather than answering an unanswerable question. The core needs only the Python 3.10+ standard library; Postgres, Redis, Docker, and GPU support are opt-in extras.

Features

• Hybrid recall: BM25F field weighting with Porter stemming and bigram phrase matching, optional vector search over a local ONNX embedder or Pinecone, and RRF fusion
• Intent routing: nine query types — WHY, WHEN, ENTITY, WHAT, HOW, LIST, VERIFY, COMPARE, TRACE — each with their own limits, expansion settings, and graph depth
• Graph recall: two-hop cross-reference boosting surfaces blocks connected structurally rather than by shared keywords
• Governed writes: propose → review → approve_apply with receipts, diffs, snapshots, and rollback, and no auto-approve
• Integrity scanning: contradiction, drift, dead-decision, and orphan-task detection across 74 or more structural rules
• Deterministic abstention: a five-signal confidence gate returns no answer instead of an unsupported one, without calling a model
• Auto-capture: session-end hooks and transcript scanning extract decision and task language into a signals file, never into the source of truth
• Tool-output offload: mm tool-run stores a command's full output out of context and returns a handle and a bounded summary, recallable later with mm tool-recall
• Multi-agent namespaces: workspace-level and per-agent private namespaces with JSON ACLs, plus a shared fact ledger with dedup and a review gate
• MCP surface: 102 tools and 8 read-only resources over stdio or HTTP, with fail-closed bearer-token auth on HTTP
• Local-first storage: Markdown files on disk with no cloud calls or telemetry; Postgres and encrypted backends are optional
• Durability: journal-based write-ahead logging, tar.gz backups, JSONL export, and cross-platform advisory locking