@more-md/mcp-memory

Portable long-term memory for AI assistants, as an MCP server.

Connect it once and your assistant remembers things across chats. The same memory follows you from Claude to Cursor to any other MCP-capable client. You hold the key, a short PIN. The memory is encrypted at rest and lives at a URL you control, mem.more.md.

Why an MCP server

Telling a chat model to "fetch this URL and use it as your memory" looks exactly like a prompt-injection attack: a web page instructing the assistant to persist your data somewhere. Well-aligned models refuse it, and they should.

An MCP server works the other way around:

• You add this server to your assistant's config, with your memory URL and PIN. You set that connection up by hand, and that is the authorization.
• The assistant calls typed tools (remember, recall). It never treats a fetched web page as instructions.
• Every call runs over HTTPS and is authed with your PIN. Your data goes only to the memory URL you configured.

Get your memory URL and PIN

Visit mem.more.md, or ask any assistant that already minted a session for you. You get a URL like https://mem.more.md/AbC12xyz and a 4-digit PIN. Keep the pair. That is your portable memory.

Connect it

Claude Desktop, Claude Code, Cursor, and most MCP clients

Add this to your MCP config (claude_desktop_config.json, .cursor/mcp.json, .mcp.json). This uses the published npm package:

{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": ["-y", "@more-md/mcp-memory"],
      "env": {
        "MEM_URL": "https://mem.more.md/AbC12xyz",
        "MEM_PIN": "1234"
      }
    }
  }
}

To run from a local checkout instead:

{
  "mcpServers": {
    "memory": {
      "command": "node",
      "args": ["/absolute/path/to/packages/mcp-memory/dist/index.js"],
      "env": { "MEM_URL": "https://mem.more.md/AbC12xyz", "MEM_PIN": "1234" }
    }
  }
}

Restart the client. A memory server shows up with three tools.

Tools

| Tool | What it does | |------|--------------| | remember | Save a memory (title, body, memory_type: episodic or semantic, importance 1 to 10). | | recall | Search your memory by keyword or meaning before answering about the past. | | recall_day | Read everything saved on a given calendar day. |

The assistant decides when to call these. A good one saves durable facts (your name, preferences, projects) as semantic and checks recall before answering questions about earlier context. You can always tell it to save or forget something specific.

Privacy

• Encrypted at rest, reachable only with your PIN.
• Export or delete everything anytime (/export, delete tools and endpoints).
• Open source. Nothing is stored that you didn't ask for.

Develop

npm install
npm run build      # builds dist/
npm test           # unit tests with mocked fetch

MIT © more.md