MIND-Mem is a memory layer for AI coding agents that stores decisions, tasks, and facts as plain Markdown blocks in a local workspace and serves them over the Model Context Protocol (MCP) through 107 tools. Its focus is on keeping long-running agent memory correct: besides storing and retrieving, it detects contradictions between decisions, drift from informal choices that were never formalized, dead decisions nobody references, and orphan tasks.
Writes are governed. Nothing reaches the source of truth without a propose, review, and approve-apply step, and every applied change is logged with a timestamp, receipt, and diff in a hash-anchored audit chain. Recall is deterministic: with the query, admitted corpus, configuration, scoring date, and execution providers held constant, the canonical audit encoding (Q16.16 fixed point) is byte-identical across replays, while ranking scores remain ordinary floating-point values.
Retrieval combines BM25F keyword search, optional vector search, and reciprocal rank fusion, with RM3 query expansion, a nine-type intent router, two-hop graph boosting across cross-referenced blocks, an optional cross-encoder reranker, and a pre-LLM abstention gate for unanswerable questions. The core runs on the Python standard library alone, with Postgres, Redis, Docker, and GPU support as opt-in extras.
One workspace can be shared across MCP clients such as Claude Code, Cursor, Windsurf, Codex CLI, Gemini CLI, and Zed, plus OpenClaw through hooks, so memory accumulates across tools. A comparison table positions it against Mem0, Letta, and Zep on locality, governance, and contradiction detection.
Results on the full LoCoMo benchmark (10 conversations, 1,986 questions, external LLM as answerer and judge):
| Category | N | Accuracy (score ≥50) | Mean score |
|---|---|---|---|
| Overall | 1986 | 73.8% | 70.5 |
| Adversarial | 446 | 92.4% | 87.2 |
| Single-hop | 282 | 80.9% | 68.7 |
| Open-domain | 841 | 71.2% | 70.3 |
| Temporal | 96 | 66.7% | 65.9 |
| Multi-hop | 321 | 50.5% | 51.1 |
Features
- Governed writes: graduated detect-only, propose, and enforce modes, with snapshots, receipts, diffs, and automatic rollback on validation failure
- Integrity scanning: detects contradictions, drift, dead decisions, orphan tasks, and regressions with 74+ structural validation rules
- Hybrid recall: BM25F plus optional vector search fused with reciprocal rank fusion, RM3 query expansion, and an intent router
- Graph-based recall: two-hop cross-reference boosting surfaces related blocks that share no keywords with the query
- Deterministic abstention: a five-signal confidence gate declines unanswerable questions before an LLM answers
- Auto-capture: session-end hooks and transcript scans detect decision and task language and write candidate signals for review, never directly to the source of truth
- Multi-agent namespaces: per-agent private namespaces with JSON access control and a shared fact ledger with a review gate
- Tool-output offload: stores long test or build output out of context and returns a compact handle and summary for later recall
- Compiled truth pages: per-entity pages with the current understanding on top and a timestamped evidence trail checked for contradictions
- Dream cycle: scheduled background enrichment that finds missing cross-references, broken citations, and redundant entries and proposes repairs
- Audit trail: every applied proposal is hash-anchored, with byte-identical replay of the audit encoding under held-constant inputs
- Storage backends: plain Markdown by default, with Postgres and encrypted storage options
