#MCP Integration
Memory Engine connects AI agents to persistent memory through the Model Context
Protocol (MCP). An MCP client starts me mcp
as a local stdio process; that process sends authenticated requests to Memory
Engine on the agent's behalf.
AI coding agent <-> me mcp (stdio) <-> Memory Engine (HTTPS)
The agent calls tools. It does not need to handle a Memory Engine credential directly.
#Managed Harness Setup
For Claude Code, OpenCode, or Codex CLI, use the managed integration path:
me init
Run it from the repository you want to configure. It enables MCP for detected harnesses and offers to install any missing integrations. Managed integrations are user-global, but MCP availability is controlled by the machine-local profile that applies to the harness directory.
If the matched profile disables MCP or does not select the harness, its managed
MCP process starts without Memory Engine tools. Use me doctor --harness <name>
to inspect the effective policy.
Installation and policy configuration are separate operations. See Harness
Integrations for the lifecycle, provider-specific
requirements, and ~/.config/me/config.yaml.
#Authentication
me mcp authenticates in this order:
--api-key <key>ME_API_KEY- Your stored
me loginsession
A local developer installation normally uses the stored session. For a
headless, unattended, or least-privilege environment, use a personal access
token or service-account key through ME_API_KEY. Memory Engine never persists
API keys for you.
The server URL resolves from --server, then ME_SERVER, then the saved
default server, then https://api.memory.build.
#Space Modes
An explicit --space or ME_SPACE creates a locked MCP server. Its memory
tools use that one space and do not expose a space parameter.
Without either selector, manual me mcp starts in multi-space mode.
me_space_list is available and memory tools require a space argument. A
stored active space does not lock a manual MCP server.
Selecting a space never grants access. Every operation remains constrained by the authenticated principal's membership and tree grants.
#Manual Stdio Setup
Any MCP client that supports stdio can start Memory Engine directly:
me mcp
Give the client a Memory Engine session through its normal process environment,
or allow it to forward ME_API_KEY, ME_SERVER, and optionally ME_SPACE.
For example, a client configuration can launch me mcp and allowlist those
environment variable names rather than storing a raw key in its config file.
Use a locked manual server only when every request should stay in one space:
me mcp --space <space-slug>
For multi-space use, call me_space_list first and pass the returned slug to
each memory tool.
#Provider Notes
#Claude Code
Install with me claude install, then restart Claude Code. The managed plugin
supplies MCP and capture plumbing; me init decides when they are active.
#OpenCode
Install with me opencode install, then restart OpenCode. It writes a managed
MCP entry and generated plugin without credentials or runtime targeting.
#Codex CLI
Install with me codex install, then approve the installed hooks through
/hooks. Codex forwards ME_API_KEY, ME_SERVER, and ME_SPACE to the
managed MCP process when they exist in Codex's environment. Restart Codex after
changing those variables.
See me claude, me opencode, and
me codex for provider details.
#Available Tools
Once connected, an agent can inspect its context, store and search memories, manage trees, and import or export records. See the MCP Tool Reference for every tool and MCP Agent Instructions for recommended agent behavior.
#Troubleshooting
- Confirm
meis on the MCP client'sPATH. - Run
me statusto check local credentials. - Run
me doctor --harness <name>for a managed integration. - Restart the harness after changing its installation or environment.
- For Codex, confirm its hooks are approved through
/hooks.