Titan Memory MCP Server

A MCP server built with a three-tier memory architecture that handles storage as follows:

Short-term memory: Holds the immediate conversational context in RAM.
Long-term memory: Persists core patterns and knowledge over time. This state is saved automatically.
Meta memory: Keeps higher-level abstractions that support context-aware responses.

📦 Installation

CHECK OUT docs/guides/how-to.md for more information on how to install and run the server.

🚀 Quick Start

Basic Installation (uses default memory path):

npx -y @smithery/cli@latest run @henryhawke/mcp-titan

With Custom Memory Path:

npx -y @smithery/cli@latest run @henryhawke/mcp-titan --config '{
  "memoryPath": "/path/to/your/memory/directory"
}'

The server will automatically:

Initialize in the specified directory (or default location)
Maintain persistent memory state
Save model weights and configuration
Learn from interactions

📂 Memory Storage

By default, the server stores memory files in:

Windows: %APPDATA%\.mcp-titan
MacOS/Linux: ~/.mcp-titan

You can customize the storage location using the memoryPath configuration:

# Example with all configuration options
npx -y @smithery/cli@latest run @henryhawke/mcp-titan --config '{
  "port": 3000,
  "memoryPath": "/custom/path/to/memory",
  "inputDim": 768,
  "outputDim": 768
}'

The following files will be created in the memory directory:

memory.json: Current memory state
model.json: Model architecture
weights/: Model weights directory

Example usage

Usage Example:

const model = new TitanMemoryModel({
  memorySlots: 10000,
  transformerLayers: 8,
});

// Store semantic memory
await model.storeMemory("User prefers dark mode and large text");

// Recall relevant memories
const results = await model.recallMemory("interface preferences", 3);
results.forEach((memory) => console.log(memory.arraySync()));

// Continuous learning
model.trainStep(
  wrapTensor(currentInput),
  wrapTensor(targetOutput),
  model.getMemoryState()
);

🤖 LLM Integration

To integrate with your LLM:

Copy the contents of docs/llm-system-prompt.md into your LLM's system prompt
The LLM will automatically:
- Use the memory system for every interaction
- Learn from conversations
- Provide context-aware responses
- Maintain persistent knowledge

🔄 Automatic Features

Self-initialization
WebSocket and stdio transport support
Automatic state persistence
Real-time memory updates
Error recovery and reconnection
Resource cleanup

🧠 Memory Architecture

Three-tier memory system:

Short-term memory for immediate context
Long-term memory for persistent patterns
Meta memory for high-level abstractions

🛠️ Configuration Options

Option	Description	Default
`port`	HTTP/WebSocket port	`0` (disabled)
`memoryPath`	Custom memory storage location	`~/.mcp-titan`
`inputDim`	Size of input vectors	`768`
`outputDim`	Size of memory state	`768`

📚 Technical Details

Built with TensorFlow.js
WebSocket and stdio transport support
Automatic tensor cleanup
Type-safe implementation
Memory-efficient design

🔒 Security Considerations

When using a custom memory path:

Ensure the directory has appropriate permissions
Use a secure location not accessible to other users
Consider encrypting sensitive memory data
Backup memory files regularly

📝 License

MIT License - feel free to use and modify!

🙏 Acknowledgments

Built with Model Context Protocol
Uses TensorFlow.js
Inspired by synthience/mcp-titan-cognitive-memory