MCP Overview

MemMesh exposes its memory engine as an MCP server, so any MCP-capable client — Claude Code, Cursor, Windsurf, Codex, or your own — can read and write memory through a standard tool interface.

Transport

Two transports, same tools. The local stdio path needs no API key and keeps memory on your machine. The hosted HTTP path connects any MCP client to a brain on the hosted platform (app.memmesh.ai) with a bearer token — it’s live (see below).

Local (stdio) — recommended, no key

The installer wires your editor to launch the thinkfleet-memory binary directly as a child process. The config is a command, not a URL — no token, no Authorization header:

local stdio config

{
  "mcpServers": {
    "thinkfleet-memory": {
      "command": "/path/to/thinkfleet-memory",
      "args": ["--db", "/Users/you/.thinkfleet-memory/memory.db", "mcp"]
    }
  }
}

Memory is stored in SQLite at ~/.thinkfleet-memory/memory.db on the Free tier. See the Quickstart to install, and Installer & doctor for all options.

Hosted (HTTP)

Connect any MCP client to a brain on the hosted platform. Point it at the project’s MCP endpoint and authenticate with the project’s MCP server token (found in your dashboard — an API key does not carry MCP scope):

hosted HTTP config

{
  "mcpServers": {
    "memmesh": {
      "type": "http",
      "url": "https://app.memmesh.ai/v1/projects/YOUR_PROJECT_ID/mcp-server/http",
      "headers": { "Authorization": "Bearer YOUR_MCP_TOKEN" }
    }
  }
}

Mesh Router. One MeshKey can fan a single memory_search across every brain you’re entitled to, via https://app.memmesh.ai/v1/brains/mcp-server/http. Issue a MeshKey from your dashboard and use it as the bearer token.

Supported clients

The installer knows how to write the right config for each client:

Client	`--tool`	Config file
Claude Code	`claude-code`	`~/.claude.json` (`mcpServers` map)
Cursor	`cursor`	`~/.cursor/mcp.json` (`mcpServers` map)
Windsurf	`windsurf`	`~/.codeium/windsurf/mcp_config.json` (`mcpServers` map)
Codex CLI	`codex`	`~/.codex/config.toml` (`mcp_servers` table)

Run the installer with --dry-run to preview the exact entry for any tool without writing files, then hand-place it into a client not auto-detected.

ChatGPT and other hosted/web clients connect over the hosted HTTP path above rather than the installer: add a custom MCP connector pointing at https://app.memmesh.ai/v1/projects/YOUR_PROJECT_ID/mcp-server/http with your MCP token. Since these clients have no auto-observe hook, add to the connector’s instructions: “call memory_search before answering and memory_observe with the user’s message after each turn.”

Tools

The engine exposes a rich set of memory tools over MCP. The primary one is memory_observe (feed it raw text; the engine decides what to keep); the rest cover explicit writes, search/recall, prediction, graph reasoning, and governance. See the full MCP tool reference.

The exact tool set depends on your engine version. Always trust the tool list your connected client reports.