Memory MCP Server

Overview

Memory is one of the official Model Context Protocol reference servers. It implements persistent memory as a local knowledge graph, so a model can remember facts about a user or project across separate conversations instead of starting from zero every time.

How it works

1. Entities are the graph's nodes: a name, an entity type (person, organization, event...) and a list of observations.
2. Relations are directed, active-voice links between entities (e.g. John_Smith works_at Anthropic).
3. Observations are atomic facts attached to an entity, added or removed independently of the entity itself.
4. Tools cover creating/deleting entities, relations and observations, reading the whole graph, and searching or opening specific nodes by name. The graph is also exposed as a readable MCP Resource (memory://knowledge-graph) that emits update notifications on mutation.
5. Storage defaults to a local memory.jsonl file, configurable via the MEMORY_FILE_PATH environment variable.

Examples

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

Installation

You need Node.js 18+.

1. Add the server to your MCP client's config using the block in the sidebar.
2. Restart your client.
3. Optional: set MEMORY_FILE_PATH to control where the knowledge graph is stored (it defaults to a local memory.jsonl file).
4. Ask the agent to remember a fact, then start a new conversation and ask it to recall it, to confirm persistence works.