Llmem MCP

LLMem MCP Server

For agents that support the Model Context Protocol, LLMem can be registered as an MCP server. This gives you native search_memories and save_memory tools instead of making HTTP calls.

Installation

Option 1: npx (no install)

Add to your agent's MCP configuration:

Claude Code (~/.claude/settings.json):

{
  "mcpServers": {
    "llmem": {
      "command": "npx",
      "args": ["-y", "@llmem/mcp-server"],
      "env": {
        "LLMEM_API_KEY": "your-key-here"
      }
    }
  }
}

Cursor (.cursor/mcp.json):

{
  "mcpServers": {
    "llmem": {
      "command": "npx",
      "args": ["-y", "@llmem/mcp-server"],
      "env": {
        "LLMEM_API_KEY": "your-key-here"
      }
    }
  }
}

Option 2: Global install

npm install -g @llmem/mcp-server

Then reference llmem-mcp-server as the command instead of npx -y @llmem/mcp-server.

Available Tools

Once registered, the MCP server exposes these tools:

search_memories

Search the user's memory store. Call this at the start of every conversation and whenever the topic changes.

Parameters:

• query (string, required): Natural language question about the user. Example: "what do I know about this user's coding preferences"
• namespace (string, optional): Filter to a specific namespace like "preferences" or "projects"
• limit (number, optional): Max results to return. Default: 20

Returns: JSON with memories, faded_memories, and questions_to_consider arrays.

save_memory

Save a new fact about the user. Call this whenever the user shares preferences, decisions, corrections, or personal context.

Parameters:

• namespace (string, required): Category — identity, preferences, work, projects, interests, health, family, goals, opinions, location, style
• key (string, required): Short descriptive label — full_name, current_project, favorite_editor
• value (string, required): The actual information, with temporal context when relevant

Returns: Confirmation with the saved memory's ID.

Behavior Guidelines

The same rules from the llmem-memory skill apply:

1. Search before every response — recall the user's context first.
2. Weave memories naturally — don't list them back to the user.
3. Save proactively — when the user shares something about themselves, save it.
4. Confirm saves — briefly tell the user what was stored: "Saved: preferences/editor = neovim"
5. Surface contradictions — if questions_to_consider highlights an evolution, ask about it.
6. Prefer recent memories — when old and new conflict, the new one wins.
7. Never store the API key — it's in the environment, not in memory.

Environment Variables

Source: Sequoia-Port/llmem-skills — distributed by TomeVault.