Memory MCP Server

Every new chat starts without what the model learned in the previous one. The Memory server gives it a local knowledge graph to write to: entities (a person, a project, a company), directed relations between them such as works_at, and short observations attached to each entity. In a later session the model searches the graph and reads back what it stored.

Setup

npx -y @modelcontextprotocol/server-memory runs over stdio with no key. Set MEMORY_FILE_PATH to choose where the JSONL file lives; otherwise memory.jsonl is written next to the installed package code, which under npx is a cache folder. With Docker, mcp/memory keeps data in a named volume (-v claude-memory:/app/dist).

What gets remembered depends on your instructions. The README includes a sample system prompt that makes the model read memory at the start of each chat and record identity, preferences, goals and relationships.

What it exposes

• create_entities, create_relations, add_observations
• delete_entities (removes their relations too), delete_observations, delete_relations
• read_graph, search_nodes, open_nodes
• The memory://knowledge-graph resource, with change notifications

Limits

• search_nodes is a case-insensitive text match over names, types and observations; there is no semantic search.
• The README calls it a basic implementation, and all data sits in one local file.

Suited to personal assistants in Claude Desktop that should remember who you are and what you work on across chats.