How do I set up the official MCP memory server (server-memory) in Claude Desktop?

Short answer

Add the server to Claude Desktop's mcpServers config with npx -y @modelcontextprotocol/server-memory (Node.js 24 or later) or the mcp/memory Docker image, then restart Claude. Set MEMORY_FILE_PATH to an absolute path you own, such as ~/memory.jsonl, because the default file sits inside the package folder and can disappear on an update.

Setup in five steps

  1. Install Node.js 24 or later if you use npx. The README lists that as the requirement. Docker users can skip this.
  2. Open the Claude Desktop config file: ~/Library/Application Support/Claude/claude_desktop_config.json on macOS or %APPDATA%\Claude\claude_desktop_config.json on Windows. In Claude Desktop you can also reach it through Settings, Developer, Edit Config. Merge the memory entry into any existing mcpServers object instead of replacing it.
  3. Add the server under mcpServers:
{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-memory"],
      "env": { "MEMORY_FILE_PATH": "/Users/you/memory.jsonl" }
    }
  }
}

On Windows, use "command": "cmd" with args ["/c", "npx", "-y", "@modelcontextprotocol/server-memory"] and a Windows path such as "%USERPROFILE%\\memory.jsonl" (backslashes doubled in JSON).

  1. Or use Docker instead of npx:
"memory": {
  "command": "docker",
  "args": ["run", "-i", "-v", "claude-memory:/app/dist", "--rm", "mcp/memory"]
}

Note that this volume sits over the image’s code folder, so it holds the server code next to memory.jsonl. A newer image does not reach a volume that is already filled; the README describes how to copy the new code into the volume when you upgrade, keeping your graph.

  1. Quit Claude Desktop completely and reopen it. Closing the window is not enough; quit it from the menu bar or system tray. Then test: ask Claude to store a test fact, start a fresh chat and ask it to retrieve that fact explicitly. Open the memory file to confirm the line was written. If a call fails or the tools do not appear, check the MCP logs (~/Library/Logs/Claude/ on macOS, %APPDATA%\Claude\logs\ on Windows).

The one setting people miss: MEMORY_FILE_PATH

Without it, the server writes memory.jsonl into its own folder. Under npx that folder is inside node_modules, and the README warns that a reinstall, an update or clearing the cache can delete your graph. It logs a warning on startup when the path is inside node_modules. Set an absolute path you own. A leading ~ expands to your home folder; relative paths resolve against the server folder, and the parent folder must exist.

What you get

Tool What it does
create_entities Adds people, projects, companies as nodes
create_relations Links two entities, e.g. “Anna works_on Project X”
add_observations Adds facts to an entity
delete_entities / delete_observations / delete_relations Removes data
read_graph Returns the whole graph
search_nodes Searches names, types and observations
open_nodes Fetches specific entities by name

There is also a resource, memory://knowledge-graph, that mirrors the whole graph.

Make Claude actually use it

Installing the server does not make Claude check it. The client must retrieve memory and put it in context, for example through search_nodes, open_nodes, read_graph or the knowledge-graph resource. Add a line to your Project instructions or system prompt, for example: “At the start of each conversation, search memory for relevant entities. When you learn a lasting fact about me or my work, store it.” Then test it in a fresh chat.

The misunderstanding: this is not team memory

server-memory is a teaching example. It is local, it runs per user, it has no accounts, roles or audit trail, and the repository says it is not production-ready. Several Claude windows on the same machine can write to one file (the server uses a lock file), and other local MCP clients can share the graph when configured with the same MEMORY_FILE_PATH. But your colleagues’ Claude, and any cloud tool that cannot reach the file, see nothing, and there are no team accounts or permissions.

For the concept behind memory servers, see what an MCP memory server is. For a hosted alternative, compare the Mem0 MCP server.

Where Bombos fits

server-memory is a good way to see how MCP memory feels on your own machine. Bombos is the step after that: a plugin that gives every AI tool your team uses, from Claude Desktop to ChatGPT, Copilot and Cursor, one shared memory with team knowledge, learned rules and what already sits in mail, documents and repos, and that shows what each answer rests on. You can start alone with your own agents, in our cloud or on your own server.

Related questions

Where does server-memory store its data?

In a JSONL file. The default is memory.jsonl in the server's own folder; with MEMORY_FILE_PATH you point it to any absolute path. The parent folder must already exist.

Can I add server-memory as a custom connector on claude.ai?

Not as is. Custom connectors on claude.ai need a remote server reachable over HTTPS from the internet. server-memory is a local stdio server meant for clients like Claude Desktop that start it on your machine.

Can two people share the same memory.jsonl?

The README says locking relies on exclusive file creation and that a network file system without it is not safe for sharing. It is built for one machine, not for a team.

Is server-memory production-ready?

The MCP servers repository calls its servers reference implementations, meant to show how MCP works, and says they are not intended as production-ready solutions.

Sources

  1. Knowledge Graph Memory Server README
  2. Model Context Protocol servers repository
  3. Connect to local MCP servers (MCP documentation)
  4. Get started with custom connectors using remote MCP (Claude Help Center)
The first half hour · free

What does your AI forget?

  1. 01You tell us which AI you use and where it keeps starting over.
  2. 02We show what a shared memory changes there, on your own examples.
  3. 03You keep the plan, with or without Bombos.
Piet Baudoin
Piet BaudoinYou talk to Piet, from the Bombos teamDutch team from Utrecht and Nijmegen · data in Europe
We reply the same working day.