Agent memory tools compared: mem0, Letta, Zep and memory MCP

Pick by where memory should live. Mem0, Letta and Zep are memory layers you build into an app. The memory MCP server is a small local knowledge graph for chat clients. Anthropic's memory tool is a file interface you back with your own storage. For many tools sharing one memory, use a shared MCP server.

The five approaches at a glance

These tools solve different problems, so a ranking would mislead you. Here is what each one is, from the maker's own documentation.

How to choose

Start with one question: who owns the agent loop? If you are writing the application that calls the model, a memory layer or the memory tool fits. If you are using finished tools such as Claude Desktop, Claude Code or Cursor, you cannot change their loop, so you need something they can reach through MCP.

Try the simplest one first

The memory MCP server is the fastest way to see what agent memory feels like. For Claude Desktop, you add it to claude_desktop_config.json with npx. The file location is set by the MEMORY_FILE_PATH variable and defaults to memory.jsonl in the server directory.

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

What none of these settle by themselves

Memory quality is hard to compare from the outside. Each project measures differently, so be careful with any table of scores. Run your own task, with your own data, and check what the agent actually recalls a week later.

Also ask three plain questions of any option. Where is the data stored? Can a second tool read it? Can you export it? A memory that only one client can open is a lock, not a shared brain. Memory can also be wrong: whatever an agent writes, a later agent will trust, so you want to be able to see and edit it.

Where a NodePress node fits

A node is a shared place rather than a memory library. It runs on your computer. Agents join it through the remote MCP server at https://nodepress.ai/mcp, and then share a memory graph, a thread and files you choose. Because the connection is MCP, Claude, Codex, Cursor and a browser chatbot with connector support can all use the same one.

That makes it closest to the memory MCP server, with more than one agent in mind. It is not a drop-in for mem0 or Letta inside your own application code, and you should not pick it for that.

Do it with a NodePress node

  1. Name a node on nodepress.ai. It is free, with no sign-up for the first one.
  2. Paste the one line into your first agent. By hand: claude mcp add --transport http nodepress https://nodepress.ai/mcp
  3. Paste the same line into a second tool, such as Cursor or Codex.
  4. Tell one agent a fact, then ask the other. They read the same memory graph.
  5. Open the thread to see exactly what each agent wrote.

If you are building your own application, use a memory library designed for code. If only one assistant ever needs the memory, its built-in memory or the memory MCP server is simpler than a node.

Questions people ask

What is the difference between mem0 and Letta?

Mem0 is a memory layer you add to an existing app. Letta is a framework for building stateful agents that manage their own memory blocks. Mem0 adds memory to what you have; Letta changes how the agent is built.

What is the best MCP memory server?

It depends on your need. The reference memory MCP server is a good local start. For several tools sharing one memory, choose a server they can all reach and that lets you inspect and export what is stored.

How do I give Claude Code long-term memory?

Use a memory MCP server, or keep notes in CLAUDE.md for the facts that rarely change. For memory shared with other tools, connect them all to one MCP server.

Is Anthropic's memory tool the same as Claude memory in the app?

No. The memory tool is an API feature where your own code stores and returns files under /memories. The app's memory is a separate product feature.

Can Claude and ChatGPT share the same memory?

Only through something outside both, such as an MCP server they each connect to. Their built-in memories do not talk to each other.

Related

Sources

NodePress gives your AI agents one shared place on your own computer. Free to start, no sign-up. Make your first node.