@mnemodb/mcp

@mnemodb/mcp

The MnemoDB memory engine as an MCP server. Gives any MCP client — Claude Code, Cursor, Windsurf, Claude Desktop — persistent memory stored in human-readable, git-diffable .mem.md files.

claude mcp add mnemodb -- npx -y @mnemodb/mcp@latest

Why this and not built-in memory

CLAUDE.md and Claude's memory tool already give agents memory. MnemoDB isn't a replacement — it's the structure they don't provide: typed entries with lifecycle (TTL, supersession, compaction), provenance and a trust model (entries the agent labels tool are flagged untrusted and can't supersede or retire trusted ones), portability across tools, and auditable Markdown files you can git diff. Worth it when the agent maintains memory itself over many sessions, or you work across more than one tool. Full rationale: https://github.com/mnemodb/mnemodb

Tools

Eleven memory-semantic operations — things a plain memory folder can't do:

• memory_boot — always-loaded context (preambles + pinned entries); call at session start
• memory_list — list the whole store (no query needed); browse everything
• memory_recall — ranked retrieval for a query, with provenance + untrusted flag
• memory_show — one entry in full: body, all metadata, lifecycle status
• memory_history — an entry's supersession lineage (what it replaced / what replaced it)
• memory_stats — the store's self-report: counts by type/scope/provenance, budget, staleness
• memory_remember — store a typed entry (dedup, provenance stamping, supersession)
• memory_forget — auditable soft-delete to the archive (recoverable; trust-gated)
• memory_pin — set an entry's load tier (always/auto/cold); guards against pinning untrusted content
• memory_review — stale entries, contradictions, budget status
• memory_compact — the vacuum pass (dry-run by default)

Setup — Claude Code

claude mcp add mnemodb -- npx -y @mnemodb/mcp@latest

Or in .mcp.json at your project root:

{
  "mcpServers": {
    "mnemodb": { "command": "npx", "args": ["-y", "@mnemodb/mcp@latest"] }
  }
}

The store is .memory/ in MNEMO_STORE if set, otherwise in CLAUDE_PROJECT_DIR (Claude Code sets it for the MCP servers it launches), otherwise in the working directory; pass --require-project-dir to make that last fallback an error. It is created on the first memory_remember (or up front with npx @mnemodb/cli init, optional). Then add a short instruction to CLAUDE.md telling the agent to call memory_recall / memory_remember — see https://github.com/mnemodb/mnemodb/blob/main/docs/USAGE.md

Security stance

memory_remember records src as agent (the default) or tool; it never records user. The agent labels anything it took from tool output or a web page — copied or paraphrased — as tool, which flags it untrusted on recall (spec §10) and puts it under the guards: it can't be pinned to always, and can't supersede or forget a higher-trust entry.

The label is self-reported. It makes the guards apply to honest saves; it does not stop an injection that leaves the label off. user provenance comes only from someone editing a .mem.md file directly, and the agent cannot forget those entries.