MCP Overview

Remem 0.5 exposes a full Model Context Protocol (MCP) interface for agent frameworks. The Go service mounts the MCP interface at /mcp over Streamable HTTP alongside the REST API and Pebble-backed storage. For stdio-only clients, remem-mcp is a small companion binary that forwards calls to the server — it holds no data itself.

Transport

• Streamable HTTP (/mcp on the Remem server) — for hosted agents and clients that speak MCP over HTTP directly
• stdio (via the remem-mcp binary) — for local tools (Claude Desktop, Claude Code) that only speak stdio

Tools (8)

Tool	Description
store_memory	Store a new memory
search_memories	Semantic, keyword, or hybrid search
get_memory	Retrieve a memory by ID
update_memory	Update content or metadata
delete_memory	Soft or hard delete
find_related	Traverse the relationship graph
promote_to_longterm	Promote a short-term memory to long-term
list_recent_memories	Recently created/accessed memories, filtered and sorted

See Tool Reference for full parameter tables.

Resources (3)

Resource URI	Description
memory://stats	System statistics snapshot
memory://collections/recent	Recently created/accessed memories
memory://collections/important	High-importance memories

See Resource Reference for sample responses.

Connecting

Claude Desktop / Claude Code (stdio)

Run the remem-mcp binary as a stdio bridge to the running Remem server. Set its URL and credential in the client configuration:

{
  "mcpServers": {
    "remem": {
      "command": "docker",
      "command": "remem-mcp",
      "args": ["--url", "http://localhost:4545/mcp", "--api-key", "your-secret-key"]
    }
  }
}

See Claude Desktop Setup for the full walkthrough.

Any Streamable-HTTP-capable MCP client

Point the client directly at http://localhost:4545/mcp, passing the server key as a bearer token if auth is enabled — no companion binary needed.

Design notes

Tool descriptions, parameter schemas, and return formats were designed for LLM consumption. Each tool description explains not just what the tool does but when to call it and what to do with the result. This makes Remem work well with agents that choose tools autonomously.

Every tool returns {"success": true, ...} on success or {"success": false, "error": "TOOL_ERROR", "message": "..."} on failure — errors are never thrown as JSON-RPC protocol errors, so check success in the result body.