Skip to main content
Send all Memory MCP traffic to a single endpoint. Auth identifies the caller — the path is the same for every account.

Endpoint

Authentication

Memory MCP accepts either static agent keys or OAuth Bearer tokens.

Desktop / script clients

Some clients prepend Bearer . The server accepts both forms for static keys. Generate keys on Setup — not via this endpoint.

Browser LLM clients (Claude web, ChatGPT)

Browser clients obtain tokens through OAuth 2.1 Authorization Code + PKCE: Required scope: memory:read. Access tokens are opaque and short-lived (about 1 hour). Refresh tokens are opaque, rotate on each refresh, and last about 30 days. Reusing a spent refresh token revokes the whole connection family. Alternative discovery for some MCP clients: GET /.well-known/oauth-protected-resource/v1/mcp/memory. See Client setup for ChatGPT and Claude web connector steps.

JSON-RPC methods

tools/list response

The server exposes one tool in v1:
See retrieve_memory for full parameter and response documentation.

Tool result shape

Successful and error responses use HTTP 200 with this structure:
Tool-level errors set isError: true and include structuredContent.error_code:

Example: tools/call

Session ID and MCP headers

The server may return an Mcp-Session-Id response header during MCP handshake. session_id in tool arguments is optional — if omitted, the server uses that MCP session ID for the call. An explicit session_id overrides the handshake ID (for example, one ID per agent thread). The server stores your last 5 prompts per session in Redis (30-day TTL) and may include the last 3 prior prompts in the triage step on the next call when triage is enabled. This is a lightweight server-side hint — not full conversational context.