MCP server

@agentmem/mcp is a Model Context Protocol server that exposes AgentMem as tools to any MCP-compatible agent — Claude Desktop, Claude Code, Cursor, Continue, and others.

The agent gets shared memory with one config block. No SDK code to write.

Install

Add to your MCP client config (claude_desktop_config.json, .cursor/mcp.json, etc.):


{
  "mcpServers": {
    "agentmem": {
      "command": "npx",
      "args":    ["-y", "@agentmem/mcp"],
      "env":     { "AGENTMEM_API_KEY": "sk-..." }
    }
  }
}

Get a key at agentmem-dashboard.vercel.app . Restart your client.

Tools

Name	Purpose
`memory_write`	Persist a fact to long-term memory.
`memory_search`	Hybrid retrieval (semantic + keyword + entity graph).
`memory_get`	Fetch a memory by id.
`memory_delete`	Soft-delete a memory.
`memory_check_conflicts`	Detect contradictions before a write.
`scratch_set`	Store transient working memory with optional TTL.
`scratch_get`	Read a scratchpad key.
`workflow_summarize`	Generate a Markdown brief over a workflow’s memories.

Each tool’s input schema and description is read by the model at tool-selection time, so the agent picks the right one without prompting hints.

Self-hosting

Point AGENTMEM_BASE_URL at your own AgentMem deployment:


"env": {
  "AGENTMEM_API_KEY":  "sk-...",
  "AGENTMEM_BASE_URL": "https://memory.your-domain.com/v1"
}

Transport

The server speaks stdio (the standard MCP transport for local servers). It runs as a Node process spawned by your MCP client. A remote HTTP endpoint is planned for browser-based clients.