Creativity · MCP — server

MCP Memory Server (Knowledge Graph)

The MCP Memory Server is Anthropic's reference implementation of persistent memory for LLM clients. It stores a local knowledge graph of entities, relations, and free-text observations on disk, letting Claude Desktop recall facts about you, projects, and people across sessions — without shipping private data to a cloud vector store.

MCP facts

Kind: server
Ecosystem: anthropic-mcp
Language: TypeScript / Node.js
Transports: stdio

Capabilities

Tools: create_entities, create_relations, delete_entities, add_observations
Tools: search_nodes, read_graph, open_nodes for querying the graph
Storage: plain JSONL at MEMORY_FILE_PATH — portable, diff-able, back-up-able
Pairs with a system prompt that tells the model when to write vs read memory

Install

npx -y @modelcontextprotocol/server-memory

Configuration

{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-memory"],
      "env": {
        "MEMORY_FILE_PATH": "/Users/you/.claude/memory.jsonl"
      }
    }
  }
}

Frequently asked questions

How is MCP memory different from a vector database?

It is a structured knowledge graph, not an embedding store. Entities have types and relations; observations are text attached to entities. This makes it easy to audit and edit by hand, at the cost of no semantic search out of the box.

Where is the memory stored?

In a plain JSONL file at MEMORY_FILE_PATH (default ~/.claude/memory.jsonl). Everything lives on your machine — no network calls, no third-party storage.

How do I prompt the model to use memory consistently?

Add system-prompt guidance such as 'At the start of each chat, call search_nodes for the user. Persist new facts via create_entities / add_observations.' Without that nudge the model often forgets the server exists.

Sources

modelcontextprotocol/servers — memory — accessed 2026-04-20
Model Context Protocol introduction — accessed 2026-04-20