MIND-Mem

Replayable memory for AI agents. Deterministic recall with a byte-identical audit chain across runs, machines, and substrates.

Built on the MIND substrate • Governed-write • Deterministic recall • 102 MCP tools
_{MIND Language Profile: default (full tensor stdlib + Q16.16 + heap) — see Phase 10.6}

_{Current release: v5.0.2 — the 47 modules 5.0.0 deleted as "unreachable" are restored and being wired, the reachability report now separates a recorded decision from undeclared debt, and a flag-off build is inert again — measured, not asserted — see CHANGELOG (single source of truth; per-version detail tables below may lag the changelog)}

MIND-Mem is a deterministic AI memory system: recall is a pure function of (corpus, config, scoring_instant) — pass the same three and the same ranked results come out, on any host, on any day — with a Q16.16 fixed-point audit chain — byte-identical across runs, machines, and substrates — embedded in every applied decision. (Recall scoring itself is standard floating-point; the byte-identity guarantee is the Q16.16 audit/replay chain.)

scoring_instant is a UTC date and is the honest part of that claim: recency ranking is load-bearing for a coding agent, so it is not deleted, it is named. Omit it and it resolves to today in UTC — the one clock read on the whole path, taken once at the boundary, never inside the scoring loop. Its resolved value is bound into the recall attestation, so any attested run replays exactly by passing that date back.

Built on the MIND substrate. Governed-write (propose → review → approve_apply). 102 MCP tools as the surface — but the differentiator is the substrate underneath. On the same workspace, recall is deterministic given (corpus, config, scoring_instant) — same three inputs → same ranked results — and every block and audit hash is byte-identical across every architecture mind-mem builds on — the Q16.16 audit chain. (The ranking scores themselves are standard floating-point; the byte-identity guarantee is the audit/replay chain.)

Most memory layers ship tools. That is table-stakes. MIND-Mem ships a substrate: Q16.16 fixed-point encoding in the audit-hash preimage, a governance pipeline that rejects every unreviewed write, and an audit chain where every applied proposal is hash-anchored. The scoring path itself is pure Python (mind_kernels.py): the wheel ships MIND-language kernel sources under mind/ and no compiled kernel, and the optional native libmindmem.so is built from lib/kernels.c (C99). The substrate claim is the encoding, the gate and the chain — not the kernels, which are not compiled yet. The same query on the same workspace produces the same ranked recall, every time; that recall's audit/replay chain is byte-identical whether you replay it on the same machine or a different one that pulls the same workspace. That property is what makes MIND-Mem suitable as a canonical memory layer across heterogeneous agent stacks.

If your agent runs for weeks, it will drift. MIND-Mem prevents silent drift.

MIND-Mem powers the Memory Plane of the MIND Cognitive Kernel — the deterministic AI runtime architecture.

30-Second Demo

pip install mind-mem
mind-mem-init ~/my-workspace        # Create workspace
mind-mem-recall -q "API decisions" --workspace ~/my-workspace  # Hybrid BM25F search
mind-mem-scan ~/my-workspace        # Detect drift & contradictions

Output:

[1.204] D-20260215-001 (decision) — Use async/await for all API endpoints
        decisions/DECISIONS.md:11
[1.094] D-20260210-003 (decision) — REST over GraphQL for public API
        decisions/DECISIONS.md:20

_{Current release: v5.0.2 — a restoration release. 5.0.0 deleted 47 modules (14,711 LOC) because nothing imported them; that reasoning was wrong — "nothing imports it" is evidence about wiring, never about worth — and all 47 are back, with all 43 of their test files, being wired one at a time behind default-OFF flags. Two of them were never unreachable at all: an import scan cannot see shell dispatch. The gate that produced this now reports not wired, deliberately waiting (with the condition that flips it) separately from not wired, no recorded decision. The 5.0.0 governance guarantees are unchanged: no block reaches the store without a gate receipt, and RECALL_ATTEST_v2 still binds the query to a rank-ordered digest of the served set.}

Substrate Properties

Table of Contents

• Why MIND-Mem
• Features
• Integrations are the substrate working
• Benchmark Results
• Quick Start
• Health Summary
• Commands
• Architecture
• How It Compares
• Companion Tools
• Recall
• MIND Kernels
• Auto-Capture
• Multi-Agent Memory
• Governance Modes
• Block Format
• Configuration
• MCP Server
• Security
• Troubleshooting
• Built in MIND lang
• Contributing
• License

Deep-dive docs

• docs/setup.md — install, configure, wire MCP, opt in to MIND native kernels
• docs/usage.md — every surface (MCP tools by category, mm CLI, mind-mem-verify, Python library) with worked examples
• docs/client-integrations.md — 18 AI client integrations (Claude Code, Codex, Grok Build, Vibe, Gemini, Cursor, Windsurf, aider, OpenClaw, NanoClaw, NemoClaw, Continue, Cline, Roo, Zed, Copilot, Cody, Qodo) with mm install-all auto-detection
• docs/task-frames.md — task frames + the dead-end registry: [TF-...] multi-session continuity (resume_brief, mm resume) and [DE-...] negative action-space memory, matched by a deterministic declarative overlap that warns and never blocks
• docs/review.md — mm review: batch approval for the HITL queue — pending proposals with their pre-apply diff, provenance, chain status and staleness inline, approved or rejected many at once through the governed approve_apply path, with no auto-approve at any risk level
• docs/mind-mem-4b-setup.md — download + run the star-ga/mind-mem-4b full-FT model locally (transformers, exllamav2, vLLM, llama.cpp, Ollama, MindLLM)
• docs/companion-tools.md — companion tools that complement (not compete with) mind-mem: MindLLM for deterministic + evidence-chained inference, GitNexus for code knowledge-graph
• ROADMAP.md — feature roadmap (genuinely-open items at the top; bulk of v3.2.0→v4.0.0 shipped)
• CHANGELOG.md — release notes for every published version

Why MIND-Mem

Most memory plugins store and retrieve. That's table stakes.

MIND-Mem also detects when your memory is wrong — contradictions between decisions, drift from informal choices never formalized, dead decisions nobody references, orphan tasks pointing at nothing — and offers a safe path to fix it.

Novel Contributions

MIND-Mem introduces several techniques not found in existing memory systems:

Features

Hybrid BM25+Vector Search with RRF Fusion

Thread-parallel BM25 and vector search with Reciprocal Rank Fusion (k=60). Configurable weights per signal. Vector is optional — works with just BM25 out of the box.

RM3 Dynamic Query Expansion

Pseudo-relevance feedback using JM-smoothed language models. Expands queries with top terms from initial result set. Falls back to static synonyms for adversarial queries. Zero dependencies.

9-Type Intent Router

Classifies queries into WHY, WHEN, ENTITY, WHAT, HOW, LIST, VERIFY, COMPARE, or TRACE. Each intent type maps to optimized retrieval parameters (limits, expansion settings, graph traversal depth).

A-MEM Metadata Evolution

Auto-maintained per-block metadata: access counts, importance scores (clamped to [0.8, 1.5] reranking boost), keyword evolution, and co-occurrence tracking. Importance decays with exponential recency.

Deterministic Reranking

Four-signal reranking pipeline: negation awareness (penalizes contradicting results), date proximity (Gaussian decay), 20-category taxonomy matching, and recency boosting. No ML required.

Optional Cross-Encoder

Drop-in ms-marco-MiniLM-L-6-v2 cross-encoder (80MB). Blends 0.6 * CE + 0.4 * original score. Falls back gracefully when unavailable. Enabled via config.

MIND Kernel Sources and Configuration

The mind/ directory contains 26 .mind files: 18 INI-style pipeline configurations and eight MIND-language tensor-source prototypes. The configuration files are parsed by mind_ffi.py; the source prototypes are migration work and are not a native serving backend. See docs/MIND_CONFIG_VS_MIND_LANG.md for the verified split. The pure-Python scoring logic in src/mind_mem/mind_kernels.py remains authoritative. An optional C library implements the existing native scoring ABI when a compatible library is provided.

MIC/MAP — MIND IR graph serialization

Pure-Python codec for the STARGA wire formats: mic@2 (line-oriented text, LLM-readable, git-friendly) and MIC-B (varint binary, ~4× smaller). Both encode typed dataflow graphs (symbols + types + values + output) with byte-identical round-trip. Streaming parser for bounded peak memory; optional Cython accelerator via mind-mem[accelerated] (+16/+20/+36 % on parse). Two MCP tools (mic_convert, mic_inspect) and a mm mic CLI surface it for agents and operators. See docs/mic-map.md. Note that the canonical IR per RFC 0021 is mic@1 text + mic@3 binary (see mindlang.dev/docs/mic); the mic@2/MIC-B codec mind-mem ships is the back-compat lineage.

BM25F Hybrid Recall

BM25F field-weighted scoring (k1=1.2, b=0.75) with per-field weighting (Statement: 3x, Title: 2.5x, Name: 2x, Summary: 1.5x), Porter stemming, bigram phrase matching (25% boost per hit), overlapping sentence chunking (3-sentence windows with 1-sentence overlap), domain-aware query expansion, and optional 2-hop graph-based cross-reference neighbor boosting. Zero dependencies. Fast and deterministic.

Graph-Based Recall

2-hop cross-reference neighbor boosting — when a keyword match is found, blocks that reference or are referenced by the match get boosted (1-hop: 0.3x decay, 2-hop: 0.1x decay). Surfaces related decisions, tasks, and entities that share no keywords but are structurally connected. Auto-enabled for multi-hop queries.

Vector Recall (optional)

Pluggable embedding backend — local ONNX (all-MiniLM-L6-v2, no server needed) or cloud (Pinecone). Falls back to BM25 when unavailable.

Persistent Memory

Structured, validated, append-only decisions / tasks / entities / incidents with provenance and supersede chains. Plain Markdown files — readable by humans, parseable by machines.

Immune System

Continuous integrity checking: contradictions, drift, dead decisions, orphan tasks, coverage scoring, regression detection. 74+ structural validation rules.

Safe Governance

All changes flow through graduated modes: detect_only → propose → enforce. Apply engine with snapshot, receipt, DIFF, and automatic rollback on validation failure.

Adversarial Abstention Classifier

Deterministic pre-LLM confidence gate for adversarial/verification queries. Computes confidence from entity overlap, BM25 score, speaker coverage, evidence density, and negation asymmetry. Below threshold → forces abstention without calling the LLM, preventing hallucinated answers to unanswerable questions.

Auto-Capture with Structured Extraction

Session-end hook detects decision/task language (26 patterns with confidence classification), extracts structured metadata (subject, object, tags), and writes to SIGNALS.md only. Never touches source of truth directly. All signals go through /apply.

Tool-Output Offload (v4.2.0)

A single cargo test / pytest / build run dumps 10k–50k lines into an agent's context — the biggest single token sink for coding agents. mm tool-run -- <cmd> stores the full output out-of-context (a tool_outputs sibling table; SQLite by default, reuses the Postgres connection with no new DB) and returns only a compact {handle, summary}; mm tool-recall <handle> returns the full text on demand. The summary is bounded regardless of input (a 10 MB line or 100k error lines can't blow it up), fail-safe (the full text is always stored and every truncation is explicit and counted — a failure line is never silently dropped), and deterministic (pure pattern extraction, no LLM; versioned config). See docs/tool-output-architecture.md.

Concurrency Safety

Cross-platform advisory file locking (fcntl/msvcrt/atomic create) protects all concurrent write paths. Stale lock detection with PID-based cleanup. Zero dependencies.

Compaction & GC

Automated workspace maintenance: archive completed blocks, clean up old snapshots, compact resolved signals, archive daily logs into yearly files. Configurable thresholds with dry-run mode.

Observability

Structured JSON logging (via stdlib), in-process metrics counters, and timing context managers. All scripts emit machine-parseable events. Controlled via MIND_MEM_LOG_LEVEL env var.

Multi-Agent Namespaces & ACL

Workspace-level + per-agent private namespaces with JSON-based ACL. fnmatch pattern matching for agent policies. Shared fact ledger for cross-agent propagation with dedup and review gate.

Automated Conflict Resolution

Graduated resolution pipeline: timestamp priority, confidence priority, scope specificity, manual fallback. Generates supersede proposals with integrity hashes. Human veto loop — never auto-applies without review.

Write-Ahead Log (WAL) + Backup/Restore

Crash-safe writes via journal-based WAL. Full workspace backup (tar.gz), git-friendly JSONL export, selective restore with conflict detection and path traversal protection.

Transcript JSONL Capture

Scans Claude Code transcript files for user corrections, convention discoveries, bug fix insights, and architectural decisions. 16 transcript-specific patterns with role filtering and confidence classification.

MCP Server (102 tools, 8 resources)

Full Model Context Protocol server with 102 distinct tools and 8 read-only resources (6 static + 2 templated). The server makes 102 mcp.tool(...) registrations, but the consolidated recall dispatcher intentionally shadows the base recall, so the live surface is 102 distinct tool names. Works with Claude Code, Claude Desktop, Cursor, Windsurf, and any MCP-compatible client. HTTP and stdio transports; HTTP requires bearer-token auth (fail-closed) — see Token Auth (HTTP). v3.8.11 added mic_convert_tool / mic_inspect_tool (MIC/MAP wire format); v3.9.0 added compile_truth_walkthrough, recall_with_persona, pipeline_status, and reindex_dirty; v3.11.0 added validate_block, block_lineage, and add_block_edge (deterministic quality gates + typed lineage edges).

74+ Structural Checks + 3024 Unit Tests

validate.sh checks schemas, cross-references, ID formats, status values, supersede chains, ConstraintSignatures, and more. Backed by 3024 pytest unit tests covering all core modules.

Audit Trail

Every applied proposal logged with timestamp, receipt, and DIFF. Full traceability from signal → proposal → decision.

Calibration Feedback Loop

Per-block quality tracking with Bayesian weight computation. When users provide feedback (thumbs up/down) via calibration_feedback, the system maintains a rolling quality score per block over a 30-day window. Bayesian smoothing constrains calibration weights to the 0.5-1.5 range, preventing any single block from dominating or being silenced. Calibration weights integrate directly into the BM25 + FTS5 retrieval pipeline — high-quality blocks rank higher, low-quality blocks are naturally demoted. Use calibration_stats to inspect per-block quality distributions and global calibration health. With v4.llm_noise_profile enabled (default off), it also carries an llm_reliability section: a per-provider, per-domain reliability EMA fed by report_outcome and persisted to intelligence/llm_profiles.json. Reliability is evidence for an operator reading the record — nothing on the retrieval scoring path reads it.

LLM-Guided Multi-Query Expansion

Generates semantically diverse query reformulations before search — synonym expansion, specificity shifts, temporal rephrasing, and negation variants. Combines all reformulated queries with Reciprocal Rank Fusion for broader recall without sacrificing precision. Runs locally with zero API calls.

4-Layer Search Deduplication

Post-retrieval dedup pipeline: best-chunk-per-source (keeps highest-scoring chunk from each file), cosine similarity dedup (>0.85 threshold), type diversity capping (max 3 results per block type), and per-source chunk limiting. Eliminates redundant results that waste LLM context.

LLM-Guided Smart Chunking

Content-aware chunking that splits at semantic boundaries (headers, paragraph breaks, list items, code blocks) instead of fixed character counts. Produces variable-size chunks with overlap for continuity. Supports markdown, code, and prose with format-specific splitting rules.

Compiled Truth Pages

Per-entity knowledge compilation: current-best-understanding on top, timestamped evidence trail below. Contradiction detection across evidence entries with automatic flagging. Entities accumulate knowledge from all sessions — each new evidence entry is checked against existing facts.

Dream Cycle (Autonomous Memory Enrichment)

Scheduled background enrichment: scans recent memory for missing cross-references, broken citations, orphan entities, and consolidation opportunities. Generates repair proposals for stale links, detects implicit entities not yet formalized, and compacts redundant entries. Runs during idle periods with configurable depth.

Feature Completeness Matrix

Integrations are the substrate working

Because the substrate is deterministic, integrating with 17 different CLIs produces the same answers on each. That is not a coincidence — it is the point. MIND-Mem can be the canonical memory layer across heterogeneous agent stacks precisely because recall is deterministic given (corpus, config, scoring_instant) and its audit/replay chain is byte-identical regardless of which client is asking. The 17-CLI surface is a consequence of the substrate, not a feature in itself.

Honest positioning: the integrations below are software-level — the named tool talks to MIND-Mem via the Model Context Protocol. They are not commercial-customer relationships with any vendor. Full positioning policy: docs/integrations.md.

Native MCP integration with 17 AI development tools

pip install mind-mem
mm install-all

mm install-all auto-detects every supported client on your machine and writes the appropriate config file for each. MIND-Mem speaks the Model Context Protocol — any MCP-compatible client connects with one command.

Compatible with major LLM providers

MIND-Mem's recall pipeline is provider-agnostic. Tested against Anthropic Claude (3.5 Sonnet, 4.x), OpenAI GPT (4o, 5.4), Google Gemini (2.0 Flash, 3.1 Pro), Mistral Large, and local endpoints (Ollama, vLLM, llama.cpp). Compatibility is at the API contract level — the same MIND-Mem server returns the same answers regardless of which LLM is asking.

Production usage at STARGA

MIND-Mem is the daily-driver memory layer across STARGA's active projects, including mind, mindlang.dev, mind-inference, and arch-mind. First-party, verifiable in our own commit history.

What we do not claim

• ❌ "OpenAI / Microsoft / Anthropic / Google is a customer" — false. These are software-level MCP integrations, not commercial relationships.
• ❌ "Used by N production teams outside STARGA" — we have no telemetry. PyPI download counts measure installs, not active use.

If a future integration becomes a real commercial relationship (signed contract, paid pilot, named reference), it will appear in the press release first — not in the README.

Benchmark Results

MIND-Mem's recall engine evaluated on standard long-term memory benchmarks using multiple configurations — from pure BM25 to full hybrid retrieval with neural reranking.

Needle In A Haystack (NIAH)

250/250 — 100% retrieval across all haystack sizes, burial depths, and needle types.

Provenance. The full-matrix repro package is committed at benchmarks/repro/niah/ (raw per-case rows, recomputed metrics, and a manifest pinning the commit, config, seeds and hardware — produced on a clean tree, repo_tracked_files_dirty_at_run: false). make repro-verify recomputes the 250/250 headline from those raw rows rather than trusting the manifest's summary. This is first-party evidence: nobody outside STARGA has re-run it yet and reported the same metrics.determinism.decision_fingerprint — see EVIDENCE.md row 1 for exactly what "verified" does and does not mean here.

A single fact is planted at a controlled depth within a haystack of semantically diverse filler blocks. The system must retrieve the needle in its top-5 results using only a natural-language query.

Config: Hybrid BM25 + all-MiniLM-L6-v2 + RRF (k=60) + sqlite-vec. Full details: benchmarks/NIAH.md

LoCoMo LLM-as-Judge

Same pipeline as Mem0 and Letta evaluations: retrieve context, generate answer with LLM, score against gold reference with judge LLM. Directly comparable methodology.

Canonical flagship number: the full 10-conversation (1986-question) BM25 run directly below (Overall Acc>=50 73.8%, Mean 70.5) — see docs/benchmarks.md for scope, evidence, and reproduction. All other LoCoMo tables on this page are smaller historical subsamples, kept for their per-category detail — do not treat any of them as the headline number.

v1.0.7 — Hybrid + top_k=18 (external LLM answerer + judge, conv-0 subsample, 199 of 1986 questions — 10% of the full set; no raw per-question artifact is checked into this repo for this run):

Pipeline: BM25 + Qwen3-Embedding-8B (4096d) vector search → RRF fusion (k=60) → top-18 evidence blocks → observation compression → answer → judge. A/B validated: +2.8 mean vs top_k=10 baseline.

v1.1.1 — BM25 + top_k=18 (canonical — full dataset) (external LLM answerer + judge, 10 conversations, 1986 questions):

Pipeline: BM25 + RM3 query expansion → top-18 evidence blocks → observation compression → answer → judge. Full 10-conversation benchmark with the same external LLM as both answerer and judge.

v1.0.0 — BM25-only baseline (external LLM answerer + judge, 10 conversations):

Key improvements since v1.0.0: Adversarial accuracy tripled from 36.3% to 92.4% via abstention classifier + hybrid retrieval. Overall Acc≥50 improved from 67.3% to 73.8% (+6.5pp).

Competitive Landscape (LoCoMo)

Full 10-conversation (1986-question) LoCoMo, Acc>=50 — apples-to-apples scope and metric. Canonical MIND-Mem table + evidence: docs/benchmarks.md. On this metric MIND-Mem is not the top score — Memobase and Letta report slightly higher, both on cloud infrastructure with embedding + vector-DB dependencies. MIND-Mem's differentiator is not "wins every cell"; it's being the only local-only, zero-core-dependency, governed system in the table — governance (contradiction detection, drift analysis, proposal/review/apply audit trail, byte-identical replay) is a property no other row has, and it is not measured by this benchmark at all.

¹ Third-party self-reported numbers (Letta's August 2025 analysis — see "Why Plain Files Outperform Fancy Retrieval" below). Not re-run by MIND-Mem, and not measured under a shared contract: different hardware, different judge configuration, and each system's own harness. Rows in this table are therefore indicative of scope, not a head-to-head result, and none of them — including ours — has been reproduced by the others. A genuine comparison requires every system run under one adapter contract on one box with a pinned judge and >=2 reps; that run does not exist yet, and until it does no ordering in this table should be read as a ranking.

² 66.88 is Mem0's own published LoCoMo-paper number. Mem0's separate 2026 managed platform self-reports 91.6 on LoCoMo — a different setup/judge (hosted product, not the open-paper config), not apples-to-apples with this table. Surfaced rather than omitted, per policy: never publish a comparison a skeptic could catch as cherry-picked.

LongMemEval-S

The previous headline (R@5 = 85.3) is retracted, not held — it had no committed artifact, was not reproducible after two attempts, and its own per-category rows summed to 376 under a stated N=470. It is replaced by the measurement below, which ships with per-question NDJSON so anyone can recompute it. See benchmarks/STATUS.md and benchmarks/REPORT.md.

Full eligible set (470 of 500; 30 abstention questions excluded), two reps identical per question, artifacts under docs/benchmarks/2026-09-03-longmemeval-s-full-*. Configuration: BM25F/SQLite with the vector leg OFF — one leg of the product, not the shipped hybrid. The hybrid number does not exist yet and nothing here may be read as one.

Paired over the same 470 question ids: on the official strict protocol (recall_all@5) the two are statistically indistinguishable (McNemar exact, p=0.4799); the zero-dependency baseline is better on the lenient protocol (recall_any@5, p=0.0043) and on MRR (p=0.0013). An independent audit refuted every artefact explanation — the recall caps are genuinely off in the pinned config, there is no ingest truncation, and index fragment ids never reach recall — so the deficit is ordering quality, not candidate recall. That is the work in flight; we publish the number that exists rather than the one we want.

Performance (Latency & Throughput)

Measured on a single developer workstation (commodity x86-64, warm cache, single process) against a 65-block workspace (typical personal workspace) with the SQLite FTS5 backend. Absolute latencies are hardware-dependent — the portable claim is the O(log N) scaling noted below, not the millisecond figures:

Query latency scales as O(log N) with SQLite FTS5 (vs O(corpus) for scan backend). The co-retrieval graph adds < 1ms per query. Knee cutoff and fact aggregation add negligible overhead (< 0.5ms).

Feedback-Quality -> Downstream-Success (synthetic, deterministic)

Downstream-success prediction (synthetic, deterministic): starved 0.00 -> sufficient 1.00 at matched budget. 48-episode regression gate over the v4.7.0 per-hit feedback-quality credit + v4.8.0 recall-sufficiency score; see benchmarks/REPORT.md and benchmarks/feedback_success_bench.py.

Run Benchmarks Yourself

# Retrieval-only (R@K metrics)

## Install in 3 commands

```bash
pip install mind-mem
mm install-all --force      # auto-wires every detected AI CLI
mm install-model            # downloads mind-mem-4b GGUF + imports to Ollama

Full options + Postgres setup + troubleshooting: docs/install-guide.md

python3 benchmarks/locomo_harness.py python3 benchmarks/longmemeval_harness.py

LLM-as-judge (accuracy metrics, requires API key)

python3 benchmarks/locomo_judge.py --dry-run python3 benchmarks/locomo_judge.py --answerer-model --output results.json

Hybrid retrieval with any model pair (BM25 + vector + cross-encoder)

python3 benchmarks/locomo_judge.py --hybrid --compress --answerer-model --judge-model --output results.json

Selective conversations

python3 benchmarks/locomo_harness.py --conv-ids 4,7,8

---

## Quick Start

### One-line install (recommended)

```bash
pipx install "mind-mem[mcp]"
mind-mem-mcp --help          # smoke-test

pipx keeps MIND-Mem in its own venv, exposes the mind-mem-mcp console script on PATH, and avoids polluting your system Python. If you don't have pipx, pip install --user "mind-mem[mcp]" works too.

Then wire it into every AI coding client on your machine:

git clone https://github.com/star-ga/mind-mem.git
cd mind-mem
./install.sh --all --no-install   # Already installed via pipx, just wire clients

Or do both in one shot (the installer will auto-pick pipx if available, else fall back to pip):

git clone https://github.com/star-ga/mind-mem.git
cd mind-mem
./install.sh --all

This auto-detects every AI coding client on your machine and configures MIND-Mem for all of them. Each client launches the same mind-mem-mcp binary, so all agents share one workspace. Supported clients:

Selective install:

./install.sh --claude-code --codex --gemini         # Only specific clients
./install.sh --all --workspace ~/my-project/memory  # Custom workspace path

Uninstall:

./uninstall.sh          # Remove from all clients (keeps workspace data)
./uninstall.sh --purge  # Remove everything including workspace data

Manual Setup

For manual or per-project setup:

1. Clone into your project

cd /path/to/your/project
git clone https://github.com/star-ga/mind-mem.git .mind-mem

2. Initialize workspace

python3 .mind-mem/src/mind_mem/init_workspace.py .

Creates 12 directories, 19 template files, and mind-mem.json config. Never overwrites existing files.

3. Validate

bash .mind-mem/src/mind_mem/validate.sh .
# or cross-platform:
python3 .mind-mem/src/mind_mem/validate_py.py .

Expected: 74 checks | 74 passed | 0 issues.

4. First scan

python3 .mind-mem/src/mind_mem/intel_scan.py .

Expected: 0 critical | 0 warnings on a fresh workspace.

5. Verify recall + capture

python3 .mind-mem/src/mind_mem/recall.py --query "test" --workspace .
# → No results found. (empty workspace — correct)

python3 .mind-mem/src/mind_mem/capture.py .
# → capture: no daily log for YYYY-MM-DD, nothing to scan (correct)

6. Add hooks (optional)

Option A: Claude Code hooks (recommended)

Merge into your .claude/hooks.json:

{
  "hooks": [
    {
      "event": "SessionStart",
      "command": "bash .mind-mem/hooks/session-start.sh"
    },
    {
      "event": "Stop",
      "command": "bash .mind-mem/hooks/session-end.sh"
    }
  ]
}

Option B: OpenClaw hooks (for OpenClaw 2026.2+)

cp -r .mind-mem/hooks/openclaw/mind-mem ~/.openclaw/hooks/mind-mem
openclaw hooks enable mind-mem

7. Smoke Test (optional)

bash .mind-mem/src/mind_mem/smoke_test.sh

Creates a temp workspace, runs init → validate → scan → recall → capture → pytest, then cleans up.

Health Summary

After setup, this is what a healthy workspace looks like:

$ python3 -m mind_mem.intel_scan .

mind-mem Intelligence Scan Report v2.0
Mode: detect_only

=== 1. CONTRADICTION DETECTION ===
  OK: No contradictions found among 25 signatures.

=== 2. DRIFT ANALYSIS ===
  OK: All active decisions referenced or exempt.
  INFO: Metrics: active_decisions=17, active_tasks=7, blocked=0,
        dead_decisions=0, incidents=3, decision_coverage=100%

=== 3. DECISION IMPACT GRAPH ===
  OK: Built impact graph: 11 decision(s) with edges.

=== 4. STATE SNAPSHOT ===
  OK: Snapshot saved.

=== 5. WEEKLY BRIEFING ===
  OK: Briefing generated.

TOTAL: 0 critical | 0 warnings | 16 info

Commands

Architecture

your-workspace/
├── mcp_server.py            # MCP server (FastMCP, 102 tools, 8 resources)
├── mind-mem.json             # Config
├── MEMORY.md                # Protocol rules
│
├── mind/                    # 26 .mind files: 18 INI config + 8 MIND sources
│   ├── README.md            # Source/config inventory and migration status
│   ├── bm25.mind            # MIND-language source prototype
│   ├── rrf.mind             # MIND-language source prototype
│   ├── ranking.mind         # MIND-language source prototype
│   └── recall.mind          # INI pipeline configuration example
│
├── lib/                     # Optional native C scoring backend
│   └── libmindmem.so        # Locally built from lib/kernels.c; not bundled
│
├── decisions/
│   └── DECISIONS.md         # Formal decisions [D-YYYYMMDD-###]
├── tasks/
│   └── TASKS.md             # Tasks [T-YYYYMMDD-###]
├── entities/
│   ├── projects.md          # [PRJ-###]
│   ├── people.md            # [PER-###]
│   ├── tools.md             # [TOOL-###]
│   └── incidents.md         # [INC-###]
│
├── memory/
│   ├── YYYY-MM-DD.md        # Daily logs (append-only)
│   ├── intel-state.json     # Scanner state + metrics
│   └── maint-state.json     # Maintenance state
│
├── summaries/
│   ├── weekly/              # Weekly summaries
│   └── daily/               # Daily summaries
│
├── intelligence/
│   ├── CONTRADICTIONS.md    # Detected contradictions
│   ├── DRIFT.md             # Drift detections
│   ├── SIGNALS.md           # Auto-captured signals
│   ├── IMPACT.md            # Decision impact graph
│   ├── BRIEFINGS.md         # Weekly briefings
│   ├── AUDIT.md             # Applied proposal audit trail
│   ├── SCAN_LOG.md          # Scan history
│   ├── proposed/            # Staged proposals + resolution proposals
│   │   ├── DECISIONS_PROPOSED.md
│   │   ├── TASKS_PROPOSED.md
│   │   ├── EDITS_PROPOSED.md
│   │   └── RESOLUTIONS_PROPOSED.md
│   ├── applied/             # Snapshot archives (rollback)
│   └── state/snapshots/     # State snapshots
│
├── shared/                  # Multi-agent shared namespace
│   ├── decisions/
│   ├── tasks/
│   ├── entities/
│   └── intelligence/
│       └── LEDGER.md        # Cross-agent fact ledger
│
├── agents/                  # Per-agent private namespaces
│   └── <agent-id>/
│       ├── decisions/
│       ├── tasks/
│       └── memory/
│
├── mind-mem-acl.json        # Multi-agent access control
├── .mind-mem-wal/           # Write-ahead log (crash recovery)
│
└── src/mind_mem/
    ├── mind_ffi.py          # MIND FFI bridge (ctypes)
    ├── hybrid_recall.py     # Hybrid BM25+Vector+RRF orchestrator
    ├── block_metadata.py    # A-MEM metadata evolution
    ├── cross_encoder_reranker.py  # Optional cross-encoder
    ├── intent_router.py     # 9-type intent classification (adaptive)
    ├── recall.py            # BM25F + RM3 + graph scoring engine
    ├── recall_vector.py     # Vector/embedding backends
    ├── sqlite_index.py      # FTS5 + vector + metadata index
    ├── connection_manager.py # SQLite connection pool (WAL read/write separation)
    ├── block_store.py       # BlockStore protocol + MarkdownBlockStore
    ├── corpus_registry.py   # Central corpus path registry
    ├── abstention_classifier.py  # Adversarial abstention
    ├── evidence_packer.py   # Evidence assembly and ranking
    ├── intel_scan.py        # Integrity scanner
    ├── apply_engine.py      # Proposal apply engine (delta-based snapshots)
    ├── block_parser.py      # Markdown block parser (typed)
    ├── capture.py           # Auto-capture (26 patterns)
    ├── compaction.py        # Compaction/GC/archival
    ├── mind_filelock.py     # Cross-platform advisory file locking
    ├── observability.py     # Structured JSON logging + metrics
    ├── namespaces.py        # Multi-agent namespace & ACL
    ├── conflict_resolver.py # Automated conflict resolution
    ├── backup_restore.py    # WAL + backup/restore + JSONL export
    ├── transcript_capture.py  # Transcript JSONL signal extraction
    ├── validate.sh          # Structural validator (74+ checks)
    └── validate_py.py       # Structural validator (Python, cross-platform)

How It Compares

Quick Comparison

¹ Canonical MIND-Mem LoCoMo number — see docs/benchmarks.md for scope/evidence. On this apples-to-apples metric MIND-Mem is not the top score of every system evaluated (see the Competitive Landscape table above); its differentiator is being the only local-only, zero-dependency, governed option.

² Mem0's own published LoCoMo-paper number. Mem0's separate 2026 managed platform self-reports 91.6 on a different setup/judge — not apples-to-apples with this row.

At a Glance

Full Feature Matrix

Compared against every major memory solution for AI agents (as of 2026):

The Gap MIND-Mem Fills

Every tool above does storage + retrieval. None of them answer:

• "Do any of my decisions contradict each other?"
• "Which decisions are active but nobody references anymore?"
• "Did I make a decision in chat that was never formalized?"
• "What's the downstream impact if I change this decision?"
• "Is my memory state structurally valid right now?"

MIND-Mem focuses on memory governance and integrity — the critical layer most memory systems ignore entirely.

Why Plain Files Outperform Fancy Retrieval

Letta's August 2025 analysis showed that a plain-file baseline (full conversations stored as files + agent filesystem tools) scored 74.0% on LoCoMo with gpt-4o-mini — beating Mem0's top graph variant at 68.5%. Key reasons:

• LLMs excel at tool-based retrieval. Agents can iteratively query/refine file searches better than single-shot vector retrieval that might miss subtle connections.
• Benchmarks reward recall + reasoning over storage sophistication. Strong judge LLMs handle the rest once relevant chunks are loaded.
• Overhead hurts. Specialized pipelines introduce failure modes (bad embeddings, chunking errors, stale indexes) that simple file access avoids.
• For text-heavy agentic use cases, "how well the agent manages context" > "how smart the retrieval index is."

MIND-Mem's deterministic retrieval pipeline validates these findings: 73.8% on the full 10-conversation LoCoMo suite (Acc≥50, canonical run below) with zero dependencies, no embeddings, and no vector database — 5.3pp above Mem0's top graph variant (68.5%). The key insight: treating retrieval as a reasoning pipeline (wide candidate pool → deterministic rerank → context packing) matches embedding+vector systems without any ML infrastructure. Unlike plain-file baselines, MIND-Mem adds integrity checking, governance, and agent-agnostic shared memory via MCP that no other system provides.

Companion Tools

External tools that solve an adjacent problem MIND-Mem deliberately does not solve. They are listed as complements, not competitors — MIND-Mem does not depend on any of them. License, scope, and substrate-of-record concerns make co-existence the right pattern: each one is a separate process you run alongside MIND-Mem, never a package dependency.

GitNexus answers a different question

Code structure now versus governed decision history — orthogonal, and usefully so. Install both and each shows up in your MCP client's tool list answering its own question domain, with no wiring between them:

# GitNexus — follow its own README for install + MCP registration
git clone https://github.com/h4ckf0r0day/GitNexus

# MIND-Mem (Apache-2.0, this repo)
pip install "mind-mem[all]"
mm install-all   # auto-wires MCP for Claude Code, Cursor, Windsurf, ...

Full positioning, the MindLLM quick start, and the license reasoning: docs/companion-tools.md.

Recall

Default: BM25 Hybrid

python3 -m mind_mem.recall --query "authentication" --workspace .
python3 -m mind_mem.recall --query "auth" --json --limit 5 --workspace .
python3 -m mind_mem.recall --query "deadline" --active-only --workspace .

BM25F scoring (k1=1.2, b=0.75) with per-field weighting, bigram phrase matching, overlapping sentence chunking, and query-type-aware parameter tuning. Searches across all structured files.

BM25F field weighting: Terms in Statement fields score 3x higher than terms in Context (0.5x). This naturally prioritizes core content over auxiliary metadata.

RM3 query expansion: Pseudo-relevance feedback from top-k initial results. JM-smoothed language model extracts expansion terms, interpolated with the original query at configurable alpha. Falls back to static synonyms for adversarial queries.

Adversarial abstention: Deterministic pre-LLM confidence gate. Computes confidence from entity overlap, BM25 score, speaker coverage, evidence density, and negation asymmetry. Below threshold → forces abstention.

Stemming: "queries" matches "query", "deployed" matches "deployment". Simplified Porter stemmer with zero dependencies.

Hybrid Search (BM25 + Vector + RRF)

{
  "recall": {
    "backend": "hybrid",
    "vector_enabled": true,
    "rrf_k": 60,
    "bm25_weight": 1.0,
    "vector_weight": 1.0
  }
}

Thread-parallel BM25 and vector retrieval fused via RRF: score(doc) = bm25_w / (k + bm25_rank) + vec_w / (k + vec_rank). Deduplicates by block ID. Falls back to BM25-only when vector backend is unavailable.

Graph-Based (2-hop cross-reference boost)

python3 -m mind_mem.recall --query "database" --graph --workspace .

2-hop graph traversal: 1-hop neighbors get 0.3x score boost, 2-hop get 0.1x (tagged [graph]). Surfaces structurally connected blocks via AlignsWith, Dependencies, Supersedes, Sources, and ConstraintSignature scopes. Auto-enabled for multi-hop queries.

Vector (pluggable)

{
  "recall": {
    "backend": "vector",
    "vector_enabled": true,
    "vector_model": "all-MiniLM-L6-v2",
    "onnx_backend": true
  }
}

Supports ONNX inference (local, no server) or cloud embeddings. Falls back to BM25 automatically if unavailable.

MIND Kernels

MIND-Mem ships 26 .mind files under mind/: 18 INI-style pipeline configuration files and eight MIND-language tensor sources. Configuration is parsed by load_kernel_config() in src/mind_mem/mind_ffi.py; compiler sources are migration prototypes. See the file inventory and format distinction.

Native migration status

The Python implementation remains available without mindc. An optional C library implements the existing native scoring ABI. A MIND-emitted replacement still needs compiler support, consumer ABI compatibility, numerical parity and performance validation. Source verification alone does not establish those gates. See compiler development and native bridge status for the current boundary. The 26 files ship in the wheel under <sys.prefix>/share/mind-mem/kernels/; packaging them does not execute the sources.

Compiler Source Index

The other 18 files are INI-style configurations and do not define compiler functions. See the source/configuration inventory.

Performance

FFI Bridge

The compiled .so exposes a C99-compatible ABI. Python calls via ctypes through src/mind_mem/mind_ffi.py:

from mind_ffi import get_kernel, is_available, is_protected

if is_available():
    kernel = get_kernel()
    scores = kernel.rrf_fuse_py(bm25_ranks, vec_ranks, k=60.0)
    print(f"Protected: {is_protected()}")  # True with the hardened build

Without MIND

If lib/libmindmem.so is not present, MIND-Mem uses the supported pure-Python implementations. The optional native C library is a performance path; no MIND-emitted replacement or cross-backend parity claim follows from its absence.

Auto-Capture

Session end
    ↓
capture.py scans daily log (or --scan-all for batch)
    ↓
Detects decision/task language (26 patterns, 3 confidence levels)
    ↓
Extracts structured metadata (subject, object, tags)
    ↓
Classifies confidence (high/medium/low → P1/P2/P3)
    ↓
Writes to intelligence/SIGNALS.md ONLY
    ↓
User reviews signals
    ↓
/apply promotes to DECISIONS.md or TASKS.md

Batch scanning: python3 -m mind_mem.capture . --scan-all scans the last 7 days of daily logs.

Safety guarantee: capture.py never writes to decisions/ or tasks/ directly. All signals must go through the apply engine.

Multi-Agent Memory

Namespace Setup

python3 -m mind_mem.namespaces workspace/ --init coder-1 reviewer-1

Creates shared/ (visible to all) and agents/coder-1/, agents/reviewer-1/ (private) directories with ACL config.

Access Control

{
  "default_policy": "read",
  "agents": {
    "coder-1": {"namespaces": ["shared", "agents/coder-1"], "write": ["agents/coder-1"], "read": ["shared"]},
    "reviewer-*": {"namespaces": ["shared"], "write": [], "read": ["shared"]},
    "*": {"namespaces": ["shared"], "write": [], "read": ["shared"]}
  }
}

Shared Fact Ledger

High-confidence facts proposed to shared/intelligence/LEDGER.md become visible to all agents after review. Append-only with dedup and file locking.

Conflict Resolution

python3 -m mind_mem.conflict_resolver workspace/ --analyze
python3 -m mind_mem.conflict_resolver workspace/ --propose

Graduated resolution: confidence priority > scope specificity > timestamp priority > manual fallback.

Transcript Capture

python3 -m mind_mem.transcript_capture workspace/ --transcript path/to/session.jsonl
python3 -m mind_mem.transcript_capture workspace/ --scan-recent --days 3

Scans Claude Code JSONL transcripts for user corrections, convention discoveries, and architectural decisions. 16 patterns with confidence classification.

Backup & Restore

python3 -m mind_mem.backup_restore backup workspace/ --output backup.tar.gz
python3 -m mind_mem.backup_restore export workspace/ --output export.jsonl
python3 -m mind_mem.backup_restore restore workspace/ --input backup.tar.gz
python3 -m mind_mem.backup_restore wal-replay workspace/

Governance Modes

Recommended rollout:

1. Install → run in detect_only for 7 days
2. Review scan logs → if clean, switch to propose
3. Triage proposals for 2-3 weeks → if confident, enable enforce

Block Format

All structured data uses a simple, parseable markdown format:

[D-20260213-001]
Date: 2026-02-13
Status: active
Statement: Use PostgreSQL for the user database
Tags: database, infrastructure
Rationale: Better JSON support than MySQL for our use case
ConstraintSignatures:
- id: CS-db-engine
  domain: infrastructure
  subject: database
  predicate: engine
  object: postgresql
  modality: must
  priority: 9
  scope: {projects: [PRJ-myapp]}
  evidence: Benchmarked JSON performance
  axis:
    key: database.engine
  relation: standalone
  enforcement: structural

Blocks are parsed by block_parser.py — a zero-dependency markdown parser that extracts [ID] headers and Key: Value fields into structured dicts.

Configuration

All settings in mind-mem.json (created by init_workspace.py):

{
  "version": "4.9.1",
  "auto_capture": true,
  "auto_recall": true,
  "governance_mode": "detect_only",
  "recall": {
    "backend": "bm25",
    "rrf_k": 60,
    "bm25_weight": 1.0,
    "vector_weight": 1.0,
    "vector_model": "all-MiniLM-L6-v2",
    "vector_enabled": false,
    "onnx_backend": false
  },
  "proposal_budget": {
    "per_run": 3,
    "per_day": 6,
    "backlog_limit": 30
  },
  "compaction": {
    "archive_days": 90,
    "snapshot_days": 30,
    "log_days": 180,
    "signal_days": 60
  }
}

MCP Server

MIND-Mem ships with a Model Context Protocol server that exposes memory as resources and tools to any MCP-compatible client.

Pair with mind-nerve for token-cheap routing. When your agent host loads many skills/tools/MCP servers, mind-nerve sits in front and returns only the top-K relevant to each request — typically a 95%+ reduction in skill-listing tokens. Apache-2.0 wheel, pip install mind-nerve. See star-ga/mind-nerve.

Install

pipx install "mind-mem[mcp]"   # preferred — isolated venv with mind-mem-mcp on PATH
# or
pip install --user "mind-mem[mcp]"

The [mcp] extra pulls fastmcp>=3.2.0 (the version line declared in pyproject.toml) and registers the mind-mem-mcp console script.

Automatic Setup (Recommended)

./install.sh --all

Configures all detected clients automatically. See Quick Start.

Manual Setup

For Claude Code, Claude Desktop, Cursor, Windsurf, and Gemini CLI, add to the respective JSON config under mcpServers:

{
  "mcpServers": {
    "mind-mem": {
      "command": "mind-mem-mcp",
      "args": [],
      "env": {"MIND_MEM_WORKSPACE": "/path/to/your/workspace"}
    }
  }
}

mind-mem-mcp is the console script registered by pipx install "MIND-Mem[mcp]" (or pip install --user "MIND-Mem[mcp]"). If you're running out of a source checkout instead, replace "command": "mind-mem-mcp" with "command": "python3", "args": ["/path/to/mind-mem/mcp_server.py"].

For Codex CLI (TOML format), add to ~/.codex/config.toml:

[mcp_servers.mind-mem]
command = "mind-mem-mcp"
args = []

[mcp_servers.mind-mem.env]
MIND_MEM_WORKSPACE = "/path/to/your/workspace"

For Zed, add to ~/.config/zed/settings.json under context_servers:

{
  "context_servers": {
    "mind-mem": {
      "command": {
        "path": "mind-mem-mcp",
        "args": [],
        "env": {"MIND_MEM_WORKSPACE": "/path/to/your/workspace"}
      }
    }
  }
}

Direct (stdio / HTTP)

# stdio transport (default)
MIND_MEM_WORKSPACE=/path/to/workspace mind-mem-mcp

# HTTP transport (multi-client / remote) — requires MIND_MEM_TOKEN per v3.7.0 fail-closed contract
MIND_MEM_WORKSPACE=/path/to/workspace MIND_MEM_TOKEN=$(openssl rand -hex 32) \
  mind-mem-mcp --transport http --host 127.0.0.1 --port 8765

Resources (read-only)

Tools (21)

Token Auth (HTTP)

MIND_MEM_TOKEN=your-secret mind-mem-mcp --transport http --port 8765

As of v3.7.0, HTTP authentication fails CLOSED. If neither MIND_MEM_TOKEN nor MIND_MEM_ADMIN_TOKEN is set, the server refuses to start. For local development you can opt back into the legacy behaviour, but only on a loopback bind:

MIND_MEM_ALLOW_UNAUTHENTICATED_LOCALHOST=1 \
  mind-mem-mcp --transport http --host 127.0.0.1 --port 8765 \
               --allow-unauthenticated-localhost

The flag is a no-op if the bind host isn't 127.0.0.1 / ::1 / localhost — the server still refuses to start. Production deployments should always set a token.

The standalone mm http-serve adapter also enforces route privileges when MIND_MEM_ADMIN_TOKEN is configured. Send either credential through X-MindMem-Token: the user token can access user routes, while the admin token also authenticates and may access admin routes. A user request to an admin route receives HTTP 404. Setting the admin variable to an empty or comma-only value keeps admin routes closed; leaving it unset preserves legacy single-token full access. Authentication and route authorization share one credential snapshot per request, refreshed for each request on a persistent connection, so a token rotation takes effect on the next request.

Safety Guarantees

• propose_update never writes to DECISIONS.md or TASKS.md. All proposals go to SIGNALS.md.
• approve_apply defaults to dry_run=True. Creates a snapshot before applying for rollback.
• All resources are read-only. No MCP client can mutate source of truth through resources.
• Namespace-aware. Multi-agent workspaces scope resources by agent ACL.

Security

Threat Model

No Network Calls

MIND-Mem makes zero network calls from its core. No telemetry, no phoning home, no cloud dependencies. Optional features (vector embeddings, cross-encoder) may download models on first use.

Requirements

• Python 3.10+
• No external packages — stdlib only for core functionality

Optional Dependencies

mind-mem:4b — Purpose-Trained LLM

⏳ Coming: a retrained mind-mem-4b. The weights published today are a full fine-tune of Qwen3.5-4B, trained against an earlier state of this repo. A full retrain on a newer base model is planned, generated from the current tool surface — several upcoming MIND-Mem features depend on it, because the shipped weights predate the surfaces those features expose. Until it lands, the current model remains the recommended one and everything below applies to it. The throughput figures in this section were measured on the current weights and will be restated when the new model ships. Not released yet; no date promised.

For best LLM extraction quality, use mind-mem:4b — a full fine-tune of Qwen3.5-4B on MIND-Mem's 8 extraction tasks (entity extraction, fact extraction, observation compression, contradiction detection, governance analysis, intent classification, axis-aware retrieval, LLM reranking). Empirical on RTX 3080 (Q4_K_M, 2.6GB VRAM): 104 tok/s generation, 1585 tok/s prefill.

Ollama (recommended):

# Download the GGUF from HuggingFace
wget https://huggingface.co/star-ga/mind-mem-4b/resolve/main/mind-mem-4b-Q4_K_M.gguf

# Create Ollama model
cat > Modelfile << 'EOF'
FROM ./mind-mem-4b-Q4_K_M.gguf
SYSTEM "You are mind-mem, a governance-aware memory extraction assistant."
PARAMETER temperature 0.1
PARAMETER num_ctx 8192
PARAMETER num_predict 1024
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|endoftext|>"
EOF
ollama create mind-mem:4b -f Modelfile

Then set in mind-mem.json:

{
  "extraction": {
    "enabled": true,
    "model": "mind-mem:4b",
    "backend": "ollama"
  }
}

Empirical on RTX 3080 (Q4_K_M, 2.6GB VRAM): 104 tok/s generation, 1585 tok/s prefill.

Full fine-tune (transformers, no adapter):

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("star-ga/mind-mem-4b", device_map="auto", torch_dtype="bfloat16")
tokenizer = AutoTokenizer.from_pretrained("star-ga/mind-mem-4b")

Platform Support

Troubleshooting

FAQ

No results from recall? Check that the workspace path is correct and points to an initialized workspace containing decisions, tasks, or entities. If the FTS5 index is stale or missing, run the reindex MCP tool to rebuild it.

MCP connection failed? Verify that fastmcp is installed (pip install fastmcp). Check the transport configuration in your client's MCP config (stdio vs HTTP). Ensure the MIND_MEM_WORKSPACE environment variable points to a valid workspace directory.

MIND sources or native kernels not loading? The eight MIND-language files are migration prototypes and are not required by the supported Python path. The optional native backend is the C implementation in lib/kernels.c; it is not bundled in the wheel. See mind/README.md for the current source and ABI boundary.

Index corrupt? Run the reindex MCP tool, or from the command line: python3 -m mind_mem.sqlite_index --rebuild --workspace /path/to/workspace. This drops and recreates the FTS5 index from all workspace files.

Specification

For the formal grammar, invariant rules, state machine, and atomicity guarantees, see SPEC.md.

MIND language sources

Eight files in mind/ contain MIND-language scoring prototypes; 18 additional .mind files are INI-style runtime configuration. The prototypes have not established a complete native backend, consumer ABI compatibility, numerical parity, or performance parity. The existing optional native implementation is the C library in lib/kernels.c, while the supported Python implementation remains available without a compiler or shared library.

The MIND language compiler is at github.com/star-ga/mind. The formal specification is at github.com/star-ga/mind-spec. The agent CLI being built on the same substrate is at github.com/star-ga/mind (RFC 0013, in development). Visit mindlang.dev to see the substrate that makes byte-identical replay possible.

The compiler and specification links above are the references for MIND-language syntax and semantics. A readable source prototype or compiler verification does not by itself establish native execution or byte-identical scoring output.

Contributing

Contributions welcome. Please open an issue first to discuss what you'd like to change.

See CONTRIBUTING.md for guidelines.

License

Apache 2.0 — Copyright 2026 STARGA Inc and contributors.