Agentmemory

AgentMemory

AgentMemory is a local-first, MIT-licensed memory layer for coding agents. It stores durable memory as plain Markdown files, keeps a daily work log and scratchpad, and can use qmd for local search.

The public project is intentionally standalone. It does not require an account, remote service, payment flow, or proprietary runtime.

Install

Or build from source with Bun:

Core commands

MCP server (agent-memory serve --mcp)

AgentMemory can also run as a real Model Context Protocol server over stdio, so any MCP-capable client or coding harness can call its tools directly instead of shelling out to individual CLI commands:

Speaks JSON-RPC 2.0, one message per line on stdin/stdout (initialize, tools/list, tools/call, ping). Exposes five tools, each backed by the exact same logic as the CLI command of the same shape:

Try it by hand:

Configure your MCP client to run agent-memory serve --mcp as a stdio server (exact config syntax varies by client). For a client that can't spawn a local process at all (e.g. a hosted connector like ChatGPT's), the recommended pattern is a relay: a small local client opens an outbound connection to a remote endpoint and forwards authenticated calls to this same stdio server, so memory content still never leaves your machine except as the direct result of a call you authorized. That relay is a separate, deployment-specific component (not part of this package) -- this server is the one piece every such setup should point at.

Memory lives under ~/.agent-memory by default:

Set AGENT_MEMORY_DIR to use a different directory.

Design principles

- Local first. Memory is stored on the developer's machine.
- Plain text. Markdown remains the source of truth.
- Portable. The same memory can be used from different coding-agent…