Connect Claude to Mnemosyne OS (MCP)

Your memory can be a tool your AI tools use. The @mnemosyne_os/mcp package (MIT) is a Model Context Protocol server that turns your local Mnemosyne OS install into a queryable memory layer for Claude Desktop, Claude Code, Cursor, Hermes Agent, and any other agent that speaks MCP.

Once connected, your agent can:

Query your code, decisions, architecture notes and git history with true semantic ranking.
Persist new decisions and insights, so future sessions recover them.
Resume a project exactly where you left off.
Filter by the nature of memories (ARCHITECTURE, SOURCE_CODE, GIT, BUGFIX…).

Everything stays on your machine. The MCP is a thin bridge that stores nothing itself. All data lives in your Mnemosyne OS vaults, and the agent only sees the vaults you allow.

It runs on itself

The Neural Map video on this wiki is Mnemosyne OS remembering her own development, and this very wiki is written by an agent connected through this exact MCP.

Requirements

Mnemosyne OS running, for the memory tools. The app owns your vaults and exposes the local gateway the MCP talks to. Three tools work with the app closed: the ones that read what other coding agents wrote on this machine (see below).
Node.js ≥ 18.

Claude Desktop, in one click

Download Mnemosyne-OS-MCP-2.1.0.mcpb (4.1 MB, or the newest one on the releases page), open Claude Desktop → Settings → Extensions, and drop the file into that panel. That is the whole install. The tools appear straight away, and the same panel offers three optional settings: default vault, other vaults, and the port the application listens on.

Claude Desktop shows a red unverified developer notice first. Every unsigned bundle does.

Double-clicking does nothing on a Microsoft Store install

A Store app does not register the .mcpb extension with Windows, so the file will not open on its own. Drop it into the Extensions panel instead.

Claude Desktop, by config file

If you would rather not install an extension:

Settings → Developer → Edit config (claude_desktop_config.json), add:

{
  "mcpServers": {
    "mnemosyne": {
      "command": "npx",
      "args": ["-y", "@mnemosyne_os/mcp"],
      "env": {
        "MNEMO_DEFAULT_VAULT": "DEV",
        "MNEMO_VAULTS": "DEV,PERSONAL,SOCIAL"
      }
    }
  }
}

Fully quit and relaunch Claude Desktop (from the tray icon, not just the window). The mnemosyne server appears under Settings → Developer → Local MCP Servers with the running badge.

Claude Code

Add .mcp.json at the root of your project with the same block plus one line, then reload the session:

{
  "mcpServers": {
    "mnemosyne": {
      "command": "npx",
      "args": ["-y", "@mnemosyne_os/mcp"],
      "env": {
        "MNEMO_DEFAULT_VAULT": "DEV",
        "MNEMO_VAULTS": "DEV,PERSONAL,SOCIAL",
        "CLAUDE_CODE_SESSION_ID": "${CLAUDE_CODE_SESSION_ID}"
      }
    }
  }
}

The last line hands the server your own session id. A stdio MCP server only receives the env its config declares, so without it the tools that list the agents on this machine cannot mark which line is you, and they say so in one sentence instead of guessing.

Cursor

Same block again, in .cursor/mcp.json.

Hermes Agent

Hermes ships with MCP support, so there is no extra install. The recipe lives in the package README.

What your agent gets

Twenty-nine tools, in seven families. Twenty-five arrive in a default install; the voice tools and the erasure tool are armed by hand with an environment variable. The full parameter reference is in the package README, and the short version of this table lives on mnemosyne-os.io/mcp.

Family	Tools	Needs the app open
Memory	`mnemosyne_memory_query`, `mnemosyne_memory_ask`, `mnemosyne_vault_list`, `mnemosyne_memory_ingest`, `mnemosyne_resonance_list`, `mnemosyne_dream_bridges`, `mnemosyne_spine_assignments`, `mnemosyne_git_log`, `mnemosyne_position_get`, `mnemosyne_position_update`, `mnemosyne_about`	yes
To-do and Agenda	`mnemosyne_todo_add`, `mnemosyne_todo_list`, `mnemosyne_todo_update`, `mnemosyne_todo_categories`, `mnemosyne_agenda_add`, `mnemosyne_agenda_list`, `mnemosyne_agenda_update`, `mnemosyne_agenda_remove`	on a dev install the headless daemon serves them app closed; an npm install needs the app
The other agents on this machine	`mnemosyne_agent_list`, `mnemosyne_agent_collisions`, `mnemosyne_agent_files`	no: they read the transcripts your harness already writes, and never their content
Cockpit	`mnemosyne_cockpit_update`	yes: a status card for this session on your canvas
Pheme	`mnemosyne_pheme_watch`, `mnemosyne_pheme_radar`	yes: the radar is as fresh as the last scan you ran, never fresher
Voice	`mnemosyne_voice_list`, `mnemosyne_voice_speak`, `mnemosyne_voice_status`	yes, behind `MNEMO_VOICE=1`
Erasure	`mnemosyne_memory_forget`	yes, behind `MNEMO_FORGET=1`: permanent, no trash, no copy kept

mnemosyne_agent_collisions is the one to call before a commit: it says whether another agent session is live on the same project and branch, because the git index is shared by every process in one working tree.

You stay the gate

MNEMO_VAULTS names the vaults the agent may see. Nothing else is reachable, and Maximum-protection vaults keep their guarantees. Writes go through the same governed doors as everything else. Your agent's session sees only the chronicles you allow.

The other direction

This page lets your agent read your memory. From 1.4.3 you can also feed it: point a vault at the notes your agent writes for itself, and ask your own memory what it knows. See connect your coding agent's memory.

And the same bridge reaches your To-do and your Agenda: since 1.4.4 the agent files the tasks you agreed on, and since 1.4.5 it reads them back, changes, completes or removes them, and does the same with your appointments, app closed included. See tasks from your coding agent.

Requirements​

Claude Desktop, in one click​

Claude Desktop, by config file​

Claude Code​

Cursor​

Hermes Agent​

What your agent gets​

You stay the gate​

The other direction​