Memory MCP Server

What this server does

Memory keeps a JSONL knowledge graph on disk and gives the model nine tools to read and write it. Entities have a name, a type (`person`, `project`, `event`), and a flat list of string observations. Relations connect entities directionally in active voice (`works_at`, `depends_on`). That is the entire model: nodes, edges, and atomic facts attached to nodes. Default storage is `memory.jsonl` next to the server binary; override with `MEMORY_FILE_PATH` to put it somewhere durable. The graph survives restarts because it is just a file. Two reads are worth knowing about. `search_nodes` matches against names, types, and observation content in one pass, so the model can ask 'what do I know about Acme?' and get the entity plus its relations back. `open_nodes` retrieves specific entities by name along with the edges between them. Deletions cascade: removing an entity strips its relations too. Boot it with `npx -y @modelcontextprotocol/server-memory` and you have everything above. No schema migration, no auth, no setup beyond pointing it at a file. The interesting part is the system prompt you pair it with. The upstream README ships a sample prompt that flags `Basic Identity`, `Behaviors`, `Preferences`, `Goals`, and `Relationships` as the categories to capture. Drop that into custom instructions and the model populates the graph as you talk to it. Without that nudge, the model rarely writes anything; with it, the file grows quickly. Curate the prompt to match what you actually want remembered.

Tools included

• create_entities

  Add nodes to the memory knowledge graph.

• create_relations

  Connect entities with named relations.

• add_observations

  Attach observations to an entity.

• search_nodes

  Search the knowledge graph by query.

• read_graph

  Read the entire knowledge graph.

The authoritative list comes from the server's tools/list at runtime. This catalog entry may lag the upstream until the next ToolPiper release.

Install in ToolPiper

ToolPiper handles the JSON configuration. Click Install, fill in any required tokens, and the server is ready in your chat.

Install in another client

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

Example prompts

Security notes

Local-only. No network calls, no auth. The graph is a JSONL file in plaintext, no encryption. Anything the model writes (names, emails, opinions about people) lands there as-is, so do not point `MEMORY_FILE_PATH` at a synced folder unless you have thought about who else reads it. The system prompt determines what gets captured: an aggressive 'remember everything' prompt produces a noisier file with more PII in it. Curate the prompt, not the file after the fact. The Docker installation persists the graph to a named volume (`claude-memory`); the README warns that updating the image can overwrite an older volume's `index.js`, so back up before upgrading.

Related tools