A reference MCP server that gives an assistant persistent memory by storing a local knowledge graph of entities, relations and observations.
PluginMCP serverFree
Step 1: Make sure Node.js 24 or later is installed if you run it through npx.
Step 2: Add a memory entry to claude_desktop_config.json using npx with -y @modelcontextprotocol/server-memory, or the Docker command from the README.
Step 3: Optionally set MEMORY_FILE_PATH to choose where the memory file is stored.
Step 4: Add a system prompt, for example in a Claude.ai Project, telling the model when to retrieve and record memory.
Step 5: Chat as usual; the model can then create entities and search the graph.
In a local JSONL file, memory.jsonl in the server directory by default. Set MEMORY_FILE_PATH to use another path.
Yes. Write tools use an exclusive lock file so writes from different processes take turns. This is not safe on a network filesystem that does not honour exclusive file creation.
That depends on your prompt. The README gives an example system prompt, and notes that changing it changes how often and what kinds of memories are created.
Requirements
M3 Memory MCPm3 Memory is a local-first memory layer that lets AI coding assistants keep facts, decisions and conversation history across sessions and tools. Assistants reach it through MCP, and it also has a command-line interface.
Inspeximus MCPInspeximus is tamper-evident long-term memory for AI agents, shipped as one zero-dependency Python file plus an MCP server. A corrected fact stays corrected, and erasures can be proven.Sage MCPSAGE is a governed memory layer for AI agents. Each memory is signed by an agent, admitted through a review step, limited by permissions and kept with an audit trail.