MCP Server
Memory MCP Server
Official reference implementation providing persistent memory through a local knowledge graph with entities, relations, and observations stored in JSONL format
What is Memory MCP Server?
The Memory MCP Server is a reference implementation that enables AI assistants to maintain persistent knowledge about users across conversation sessions through a local knowledge graph. The server models information as entities (nodes with names, types, and observations), relations (directed connections in active voice), and observations (atomic facts attached to entities). All data is persisted to a JSONL file (default: memory.jsonl in the server directory, configurable via MEMORY_FILE_PATH environment variable) with each mutation appending an operation line for durability. The server exposes 10 tools including create_entities, create_relations, add_observations for graph construction, delete_entities, delete_observations, delete_relations for cleanup, and read_graph, search_nodes, open_nodes for retrieval. A knowledge-graph MCP Resource at memory://knowledge-graph provides the full graph as JSON with automatic notifications/resources/updated broadcasts when mutation tools modify the graph. This enables clients subscribing to the resource to maintain synchronized views of the knowledge state.
Key capabilities
Knowledge Graph with Entities, Relations, and Observations
Models persistent memory as a directed graph where entities are typed nodes with unique names, relations are active-voice connections between entities, and observations are atomic facts, all stored in a structured JSONL format for durability.
JSONL-Based Persistence with Append-Only Operations
Writes each mutation (create, add, delete) as a new JSONL line to the memory file, providing durability and audit trail while allowing the graph to be reconstructed by replaying operations from the file on server initialization.
MCP Resource with Automatic Update Notifications
Exposes the knowledge graph as a memory://knowledge-graph Resource with MIME type application/json, broadcasting notifications/resources/updated after each mutation tool invocation so subscribed clients can refresh their views.
Search and Retrieval with Node Opening and Graph Queries
Provides search_nodes for text search across entity names, types, and observations, open_nodes for retrieving specific entities and their relations, and read_graph for full graph dump, enabling flexible knowledge retrieval patterns.
Use cases
User Preference Persistence Across Sessions
AI assistants can create entities for users, store observations about preferences, work patterns, or project context, and recall this information in future conversations without requiring users to repeat background information each time.
Project Context Tracking with Entity Relations
Developers can have AI assistants track project entities (repositories, team members, dependencies) and their relationships (works_on, depends_on, maintained_by), building a queryable graph of project knowledge that persists across coding sessions.
Conversational Context Accumulation
As users interact with AI assistants, the server accumulates observations about discussed topics, decisions made, and action items, creating a growing knowledge base that improves contextual understanding over time.
Docker Volume Persistence for Containerized Deployments
When running in Docker with a volume mount to /app/dist, the JSONL file persists across container restarts, allowing knowledge graphs to survive deployments while maintaining isolation from the host filesystem.
MCP capabilities
Tools
create_entities
create_relations
search_nodes
Resources
memory://knowledge-graph
Connection and auth
- Node.js runtime for NPX execution or Docker for containerized deployment
- Writable filesystem location for JSONL storage (configurable via MEMORY_FILE_PATH)
- MCP client supporting Resource subscriptions for live graph updates
Client configuration
Claude Desktop
{
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-memory"
]
}