Claude Code Agentic Semantic Memory System MCP

🚀 Claude Code Agentic Semantic Memory System (MCP)

A sophisticated Model Context Protocol (MCP) server that provides Claude Code with persistent semantic memory capabilities, enabling long-term information storage and retrieval across sessions. This system allows Claude to remember context, preferences, and important information between conversations.

✨ Features

• 🧠 Persistent Memory: Information survives across Claude Code sessions
• 🔍 Semantic Search: Retrieve memories based on meaning, not just keywords
• 📁 Project Namespaces: Organize memories into separate contexts/projects
• 🌐 Local Embeddings: Works offline without external API dependencies
• 🔗 Memory Relations: Create knowledge graphs with connected memories
• 🚀 Intent-Based Activation: Natural language triggers automatic tool invocation
• 📊 Full CRUD Operations: Create, read, update, and delete memories
• 🔐 Privacy-First: All data stored in your own PostgreSQL database

🏗️ Architecture

Core Components

agentic-memory/
├── src/
│   ├── index.ts           # MCP server entry point
│   ├── db/
│   │   ├── client.ts      # Database connection (Neon PostgreSQL)
│   │   └── schema.ts      # Drizzle ORM schema (memories, relations)
│   └── tools/
│       ├── createMemory.ts    # Store new memories
│       ├── searchMemory.ts    # Semantic search
│       ├── listMemories.ts    # List and filter memories
│       ├── deleteMemory.ts    # Remove memories
│       ├── updateMemory.ts    # Modify existing memories
│       ├── manageProjects.ts  # Project namespace management
│       └── simpleEmbedding.ts # Local embedding generation
├── run-server.sh          # Server startup script
└── .env                   # Environment configuration

How It Works

1. Claude Code sends requests to the MCP server via stdio
2. MCP Server processes tool requests (create, search, delete memories)
3. Embeddings are generated locally using mathematical algorithms
4. PostgreSQL (Neon) stores memories with pgvector for semantic search
5. Drizzle ORM manages database operations and migrations

🚀 Quick Start

Prerequisites

• Node.js 18+
• PostgreSQL database with pgvector extension (Neon recommended)
• Claude Code CLI installed

Installation

1. Clone or copy this MCP server to your project:

cp -r /path/to/agentic-memory /your/project/.claude/mcp-servers/

2. Install dependencies:

cd /your/project/.claude/mcp-servers/agentic-memory
npm install

3. Configure environment variables:

cp .env.example .env
# Edit .env with your database URL

4. Run database migrations:

npm run db:migrate

5. Add to Claude Code:

claude mcp add agentic-memory ./run-server.sh

6. Verify connection:

claude mcp list
# Should show: agentic-memory - ✓ Connected

💻 Usage Examples

Available MCP Tools

Using with CLAUDE.md Orchestration

The CLAUDE.md file acts as a quasi-orchestrator, providing:

• Context Instructions: High-level rules for memory usage
• Specialist Agents: Memory agents that Claude Code can invoke
• Immutability Principles: Memories should be append-only
• Quality Guidelines: Only store high-signal, reusable information

Example CLAUDE.md pattern:

# Project Context: [Your Project Name]

This project uses an advanced semantic memory system to learn and improve over time.

## Key Capabilities
- **Semantic Memory**: Long-term memory storage and retrieval
- **Specialist Agents**: Dedicated agents for memory operations

## Core Principles
- **Immutable Memories**: Once created, memories should not be altered
- **High-Signal Only**: Store only reusable, non-obvious information
- **No Secrets**: Never store API keys, passwords, or PII

🔧 Technical Details

Configuration

Environment Variables

Create .env file:

# PostgreSQL connection (required)
DATABASE_URL=postgresql://user:pass@host/db?sslmode=require

# Optional: Embedding configuration
EMBEDDING_MODEL=simple  # or 'llama' for advanced
EMBEDDING_DIMENSIONS=1536

# Optional: Default project
DEFAULT_PROJECT=default

Database Schema

The system uses two main tables:

• memories: Stores content, embeddings, metadata, and timestamps
• memory_relations: Links related memories (parent-child relationships)

Embedding Systems

Simple Embeddings (Default)

• Pros: No external dependencies, works offline, fast
• Cons: Less semantic accuracy
• Use Case: Development, restricted networks, testing

Llama CPP Embeddings (Advanced)

• Pros: Higher semantic accuracy
• Cons: Requires model download, more resources
• Use Case: Production environments with good connectivity

Agent Architecture

Current Agents

✅ Memory CRUD Agents: Create, read, update, delete operations ✅ Project Management Agents: Namespace organization ✅ Search Agent: Semantic similarity search

Recommended Additional Agents

1. Memory Relations Agent

   • Create parent-child relationships
   • Build knowledge graphs
   • Track memory evolution

2. Memory Analytics Agent

   • Usage statistics
   • Memory patterns
   • Optimization recommendations

3. Export/Import Agent

   • Backup memories
   • Transfer between projects
   • Migration tools

4. Deduplication Agent

   • Detect similar memories
   • Merge duplicates
   • Maintain consistency

🚨 Common Issues & Solutions

Issue: "Failed to connect" in MCP list

Solution: Check .env file exists and DATABASE_URL is correct

Issue: Embedding dimension mismatch

Solution: Ensure EMBEDDING_DIMENSIONS matches your database vector size

Issue: Memories not persisting across sessions

Solution: Verify database connection and that migrations have run

📝 Development

Running TypeScript directly:

npm run dev

Building for production:

npm run build

Running tests:

npm test

🔄 Replication Guide

To replicate this setup in another codebase:

1. Copy the MCP server directory
2. Set up a PostgreSQL database with pgvector
3. Configure environment variables
4. Run migrations
5. Add to Claude Code configuration
6. Create a CLAUDE.md file with memory instructions
7. Test with simple memory operations

📊 Production Considerations

• Database Scaling: Consider connection pooling for heavy usage
• Embedding Cache: Implement caching for frequently accessed memories
• Backup Strategy: Regular database backups are essential
• Security: Never store sensitive information in memories
• Monitoring: Track memory usage and search performance

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Areas for Contribution

• Improve embedding algorithms
• Add new memory management tools
• Enhance search capabilities
• Create visualization tools
• Optimize database queries
• Add support for more embedding models

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• Anthropic for Claude and the MCP protocol
• Neon for serverless PostgreSQL
• Drizzle ORM for database management
• The Claude Code community for feedback and suggestions

📧 Contact

Tristan McInnis

• GitHub: @tristan-mcinnis
• Project: claude-code-agentic-semantic-memory-system-mcp

Built for Claude Code - Enabling persistent, semantic memory across AI coding sessions.

⭐ If you find this project useful, please consider giving it a star on GitHub!