mnemon-mcp

Persistent layered memory for AI agents. Local-first. Zero-cloud. Single SQLite file.

Landing Page · npm · GitHub

Your AI agent forgets everything after each session. Mnemon fixes that.

It gives any MCP-compatible client — OpenClaw, Claude Code, Cursor, Windsurf, or your own — a structured long-term memory backed by a single SQLite database on your machine. No API keys, no cloud, no telemetry. Just npm install and your agent remembers.

Why Layered Memory?

Flat key-value stores treat "what happened yesterday" the same as "never commit without tests." That's wrong — different kinds of knowledge have different lifetimes and access patterns.

Mnemon organizes memories into four layers:

A journal entry from last Tuesday and a coding rule that never changes live in different layers — because they should.

Retrieval Quality

Retrieval is measured against a 50-case golden set on a real 797-memory bilingual (RU/EN) corpus, through the actual MCP server — not a reimplementation. Current numbers (methodology & history):

Hybrid beats both legs individually, which is the whole argument for fusing them: lexical search has the better raw recall, vector search the better ranking, and RRF keeps both instead of averaging them away.

The eval doc tracks the failures too — score drift under corpus growth, the BM25 field-weight bug the eval caught, the two cases where fusion still loses to pure lexical search, and what the golden set does not cover. Numbers you can't audit are marketing; read how these are produced.

Architecture

flowchart LR
    C["MCP client<br/>Claude Code · Cursor · …"] -- "stdio / HTTP" --> T["10 tools · 4 resources · 3 prompts"]
    T --> R["retrieval pipeline<br/>FTS5 · vector · RRF fusion"]
    T --> M["memories + supersede chains"]
    I["KB import pipeline<br/>markdown → memories"] --> M
    M -- triggers --> F["FTS5 index (stemmed EN+RU)"]
    R --> F
    R --> V["sqlite-vec (optional, BYOK)"]

One SQLite file holds memories, the FTS5 index, and the optional vector index. Writes go through transactions that keep the supersede-chain invariant; reads run the staged retrieval pipeline described under Search.

The full picture — module boundaries, write/read paths, invariants, and known limitations — is in docs/ARCHITECTURE.md. Design decisions are recorded as ADRs: SQLite+FTS5 core, hybrid RRF retrieval, synchronous driver, layered memory model.

Quick Start

Install

npm install -g mnemon-mcp

Or from source:

git clone https://github.com/nikitacometa/mnemon-memory-mcp.git
cd mnemon-memory-mcp && npm install && npm run build

Configure Your MCP Client

openclaw mcp register mnemon-mcp --command="mnemon-mcp"

{
  "mnemon-mcp": {
    "command": "mnemon-mcp"
  }
}

{
  "mcpServers": {
    "mnemon-mcp": {
      "command": "mnemon-mcp"
    }
  }
}

{
  "mcpServers": {
    "mnemon-mcp": {
      "command": "mnemon-mcp"
    }
  }
}

{
  "mnemon-mcp": {
    "command": "node",
    "args": ["/absolute/path/to/mnemon-mcp/dist/index.js"]
  }
}

Verify

echo '{"jsonrpc":"2.0","method":"tools/list","id":1}' | mnemon-mcp

You should see 10 tools in the response. The database (~/.mnemon-mcp/memory.db) is created automatically on first run.

That's it. Your agent now has persistent memory.

What It Can Do

10 MCP Tools

MCP Resources & Prompts

Resources — live data your agent can read:

Prompts — pre-built workflows:

Search

Four modes, all supporting layer / entity / scope / date / confidence filters:

FTS mode (default without embeddings) — tokenized full-text search with BM25 ranking. Multi-word queries use AND; if too few results, OR supplements with a score penalty. Progressive AND relaxation tries top-3 most specific terms before falling back to full OR.

Hybrid mode (default when embeddings configured) — combines FTS5 + vector search via Reciprocal Rank Fusion. Detects quoted entities in queries (e.g., 'Essentialism') and runs weighted sub-queries for cross-reference retrieval.

Vector mode — pure cosine similarity search over embeddings.

Exact mode — LIKE substring match for precise phrase lookups.

Scores: bm25 × (0.3 + 0.7 × importance) × decay(layer) × recency

Recency boost: 1 / (1 + daysSince / 365) — gently rewards recently created memories without penalizing old ones.

Stemming

Snowball stemmer applied at both index time and query time for English and Russian. This means "running" matches "runs", and "книги" matches "книга". Stop words are filtered from queries to improve precision.

Fact Versioning

Knowledge evolves. Mnemon doesn't delete old facts — it chains them:

v1: "Team uses React 17"  →  superseded_by: v2
v2: "Team uses React 19"  →  supersedes: v1 (active)

Search returns only the latest version. memory_inspect with include_history: true reveals the full chain. memory_delete re-activates the predecessor — nothing is lost.

Vector Search (Optional, BYOK)

Enable semantic similarity search by providing your own embedding API:

# OpenAI
MNEMON_EMBEDDING_PROVIDER=openai MNEMON_EMBEDDING_API_KEY=sk-... mnemon-mcp

# Ollama (local, free)
MNEMON_EMBEDDING_PROVIDER=ollama mnemon-mcp

This unlocks two additional search modes:

• mode: "vector" — pure cosine similarity search
• mode: "hybrid" — FTS5 + vector combined via Reciprocal Rank Fusion

Requires sqlite-vec (installed as optional dependency). New memories are embedded on add; existing ones can be backfilled.

Importing a Knowledge Base

Got a folder of Markdown files? Import them in bulk:

cp config.example.json ~/.mnemon-mcp/config.json   # edit this first
npm run import:kb -- --kb-path /path/to/your/kb     # incremental (skips unchanged files)

The config maps glob patterns to memory layers:

{
  "owner_name": "your-name",
  "extra_stop_words": [],
  "mappings": [
    {
      "glob": "journal/*.md",
      "layer": "episodic",
      "entity_type": "user",
      "entity_name": "$owner",
      "importance": 0.6,
      "split": "h2"
    },
    {
      "glob": "people/*.md",
      "layer": "semantic",
      "entity_type": "person",
      "entity_name": "from-heading",
      "importance": 0.8,
      "split": "h3"
    }
  ]
}

Config Fields

HTTP Transport

For remote or multi-client setups:

MNEMON_AUTH_TOKEN=your-secret MNEMON_HOST=0.0.0.0 MNEMON_PORT=3000 npm run start:http

Binds to 127.0.0.1 by default. Binding to any other host requires MNEMON_AUTH_TOKEN — the server refuses to expose the memory store to the network unauthenticated (override with MNEMON_ALLOW_INSECURE_HTTP=1 on a trusted network). Rate limiting (100 req/min/IP by default), opt-in CORS, 1MB body limit, timing-safe auth, graceful shutdown on SIGTERM.

Configuration Reference

Tool Reference

How It Compares

Extended competitive analysis with sources: docs/COMPETITORS.md.

Development

npm run dev        # run via tsx (no build step)
npm run build      # TypeScript → dist/
npm run lint       # eslint (flat config)
npm test           # vitest — unit + integration + MCP dispatch + HTTP transport + hybrid RRF
npm run bench      # performance benchmarks
npm run db:backup  # backup database

CI runs build + lint + tests on Node 20 and 22, then smoke-tests the compiled server over real JSON-RPC (tools/list must match the exact tool set).

Stack: TypeScript 5.9 (strict mode), better-sqlite3, @modelcontextprotocol/sdk, Snowball stemmer, Zod, vitest.

See CONTRIBUTING.md for code guidelines.

Design Principles

• Air-gapped by default — zero telemetry, ever. Out of the box nothing leaves the machine; the only component that talks to the network is the optional embedder, and only to the provider you configure (including a local Ollama).
• Single file — one SQLite database, zero ops, instant backup via file copy.
• Deterministic search — FTS5, not embeddings, is the default. Interpretable, reproducible, no GPU needed.
• Structured over flat — layers encode access patterns; superseding chains encode time.
• Minimal — 4 production dependencies. Works everywhere Node runs.
• Measured, not asserted — retrieval changes are judged against a golden set, regressions included.

License

MIT