MCP Memory Server
This server implements long-term memory capabilities for AI assistants using mem0 principles, powered by PostgreSQL with pgvector for efficient vector similarity search.
Features
- PostgreSQL with pgvector for vector similarity search
- Automatic embedding generation using BERT
- RESTful API for memory operations
- Semantic search capabilities
- Support for different types of memories (learnings, experiences, etc.)
- Tag-based memory retrieval
- Confidence scoring for memories
- Server-Sent Events (SSE) for real-time updates
- Cursor MCP protocol compatible
Prerequisites
- PostgreSQL 14+ with pgvector extension installed:
- Node.js 16+
Setup
- Install dependencies:
- Configure environment variables: Copy `.env.sample` to `.env` and adjust the values:
Example `.env` configurations:
- Initialize the database:
- Start the server:
For development with auto-reload:
Using with Cursor
Adding the MCP Server in Cursor
To add the memory server to Cursor, you need to modify your MCP configuration file located at `~/.cursor/mcp.json`. Add the following configuration to the `mcpServers` object:
Replace `/path/to/your/memory` with the actual path to your memory server installation.
For example, if you cloned the repository to `/Users/username/workspace/memory`, your configuration would look like:
The server will be automatically started by Cursor when needed. You can verify it's working by:
- Opening Cursor
- The memory server will be started automatically when Cursor launches
- You can check the server status by visiting `http://localhost:3333/mcp/v1/health`
Available MCP Endpoints
SSE Connection
- **Endpoint**: `GET /mcp/v1/sse`
- **Query Parameters**: - `subscribe`: Comma-separated list of events to subscribe to (optional)
- **Events**: - `connected`: Sent on initial connection - `memory.created`: Sent when new memories are created - `memory.updated`: Sent when existing memories are updated
Memory Operations
- **Create Memory**
- **Search Memories**
- **List Memories**
Health Check
Response Format
All API responses follow the standard MCP format:
Or for errors:
Memory Schema
- id: Unique identifier
- type: Type of memory (learning, experience, etc.)
- content: Actual memory content (JSON)
- source: Where the memory came from
- embedding: Vector representation of the content (384 dimensions)
- tags: Array of relevant tags
- confidence: Confidence score (0-1)
- createdAt: When the memory was created
- updatedAt: When the memory was last updated