Memory MCP — Persistent AI Agent Knowledge Graph

What it is

Memory MCP is a Model Context Protocol server that provides AI agents with persistent memory via a local knowledge graph. The agent can create entities (people, projects, concepts), add observations about them, and define relationships between them. This structured memory persists across sessions.

It is part of the official MCP server collection and targets developers who want their AI agents to build and maintain a knowledge base over time rather than starting fresh each conversation.

How it saves time or tokens

Without persistent memory, every AI conversation starts from zero. You re-explain your project structure, preferences, and context. Memory MCP lets the agent store and retrieve this information, reducing repeated context-setting.

The estimated token cost for this workflow is approximately 4,000 tokens. The knowledge graph format is more structured than raw text memory, making retrieval more precise.

How to use

1. Add to your MCP configuration:

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

1. Restart Claude Code or your MCP-compatible agent.

1. The agent now has tools to create entities, add observations, and query the knowledge graph:

You: Remember that our project uses PostgreSQL 16 with pgvector
Agent: I will store this in my knowledge graph...
[Creates entity: 'Project Database' with observation: 'Uses PostgreSQL 16 with pgvector']

Example

// Example knowledge graph operations the agent performs:

// Create an entity
{"tool": "create_entities", "entities": [
  {"name": "TokRepo", "entityType": "project",
   "observations": ["Nuxt 3 + Go-Zero stack", "Port 3001 frontend"]}
]}

// Create a relationship
{"tool": "create_relations", "relations": [
  {"from": "TokRepo", "to": "PostgreSQL", "relationType": "uses"}
]}

// Search memory
{"tool": "search_nodes", "query": "database"}
// Returns: TokRepo -> uses -> PostgreSQL

Related on TokRepo

• AI Memory Tools -- Compare memory solutions for AI agents
• MCP Integrations -- Database and storage MCP servers

Common pitfalls

• The knowledge graph is stored as a local JSON file. If you delete the file, all memory is lost. Back it up if the data matters.
• Memory MCP stores everything the agent decides to remember. Over time the graph can grow large. Periodically review and prune irrelevant entities.
• The agent must be explicitly asked to remember or check memory. It does not automatically store or retrieve information unless prompted.