memory-mcp: the official Anthropic memory reference — the 2-minute install for your AI agent

Short version: memory-mcp is Anthropic's official reference implementation of an MCP memory server. It's the simplest thing that could possibly work — a local file, structured entities + relations + observations, zero external dependencies. 2-minute install. If you're new to AI memory, start here. MIT licensed.

What is memory-mcp?

memory-mcp is the official Anthropic reference for MCP memory servers. It's deliberately minimal: a local file for persistence, a simple structured format (entities, relations, observations), and tools your agent calls to read and write. No external dependencies. No vector databases. No config beyond pointing your client at it.

Because it's the reference, it will keep working. Anthropic maintains it. Every MCP improvement lands here first. You can't outgrow its basic functionality — you'll only outgrow its scale or scoping model.

Who this is for

Solo founders or 2–5 person teams standing up AI memory for the first time.
Anyone who wants to prove to themselves that memory-enabled agents actually change their workflow before investing in infrastructure.
Users who work primarily through Claude Desktop or Claude Code.
People who want something they'll never have to babysit.

Skip this if

You need per-user scoping (go to mem0), multi-agent coherence across clients (go to agentmemory), or production-grade retrieval benchmarks (go to mcp-memory-service).

What problem it solves

Most memory tools assume you want something fancy. You don't. You want your agent to remember who your co-founder is, what your company does, what you decided about last quarter — without spending an afternoon reading MCP docs.

memory-mcp is the zero-decision choice. It's the memory layer Anthropic chose to showcase. Install it, forget about it, use it. The 80% case is already solved.

How to install it (plain English)

Add one config entry. In your .mcp.json:

{ "mcpServers": { "memory": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-memory"] } } }

Restart your client. First run pulls the package via npx (~20 seconds).
Start teaching. Say "Remember that my co-founder is Amara, we're building Linea, DTC skincare, 30 people." The agent uses memory-mcp's tools to persist the facts.

That's the whole install. Full walkthrough: /memory/tools/memory-mcp.

What you can do with it (for a non-technical founder)

Stop re-explaining yourself — your agent remembers who you are, who's on your team, what your priorities are.
Build an entity graph without thinking about it — "who is Alex?" queries the graph.
Scope memory per project — memory-mcp writes to a file you can point at per-project.
Inspect the memory file directly — it's on your disk. You can open it, grep it, back it up, commit it.
Migrate when you outgrow it — the file is structured; a small script migrates to mem0 / memento-mcp when you scale.

What CLO adds on top

memory-mcp gives your agent a persistent memory. Cognition CLO gives your team a retention model. Different layers. Start with memory-mcp to see what agent memory even feels like. Add CLO when you care which of your employees is retaining the knowledge your agent surfaces.

FAQ

Where is my memory stored?

Locally, on your machine, in a structured file. No cloud. You control it.

Does it work with Cursor, Aider, and other MCP clients?

Yes. Any MCP-speaking client. Same config pattern, different config file location per client.

Can I edit the memory by hand?

Yes. Open the file. It's structured but human-readable.

What happens when I outgrow it?

Move to mem0 (multi-user scoping + SDK), memento-mcp (richer graph inspection), or mcp-memory-service (coding-agent-tuned). Export, transform, import.

Is there a managed cloud version?

No — memory-mcp is local-first by design. For managed memory, mem0.ai or supermemory.ai are the common cloud picks.

Why use this over mem0?

For a solo founder experimenting: memory-mcp is a 2-minute install vs ~20 for mem0. For a multi-user product: mem0's user_id scoping is non-negotiable. Different stages, different tools.

Ready to install? Full walkthrough at /memory/tools/memory-mcp. Comparison with mem0: /blog/mem0-vs-memory-mcp. Credit to the Anthropic team — the reference keeps the ecosystem honest.