📌Overview

Purpose: This framework aims to provide a robust TypeScript-based MCP server that enables interaction with multiple LLM providers while effectively managing conversation history.

Overview: The MemGPT MCP Server is designed to facilitate seamless chatting with various LLMs, maintaining a comprehensive memory system that enhances user interaction and experience by remembering past conversations.

Key Features:

• Chat Tool: Enables users to send messages to the chosen LLM provider, supporting various options like OpenAI and Anthropic, enhancing flexibility in communication.

• Memory Management: The get_memory tool retrieves conversation history, with a configurable limit for memory retrieval, allowing for efficient management of past interactions. The clear_memory function provides an option to reset conversation history entirely.

• Flexible Provider and Model Usage: Users can easily switch between different LLM providers and models tailored for specific tasks, ensuring optimal performance and relevance in interactions. The system supports multiple models from providers like OpenAI and Anthropic, enhancing versatility.

---

MemGPT MCP Server

A TypeScript-based MCP server that implements a memory system for LLMs. It provides tools for chatting with different LLM providers while maintaining conversation history.

Features

Tools

• chat
  Send a message to the current LLM provider.

  • Takes a message parameter.
  • Supports multiple providers: OpenAI, Anthropic, OpenRouter, Ollama.

• get_memory
  Retrieve conversation history.

  • Optional limit parameter to specify number of memories to retrieve.
  • Pass limit: null for unlimited memory retrieval.
  • Returns memories in chronological order with timestamps.

• clear_memory
  Clear conversation history by removing all stored memories.

• use_provider
  Switch between different LLM providers.

  • Supports OpenAI, Anthropic, OpenRouter, and Ollama.
  • Persists provider selection.

• use_model
  Switch to a different model for the current provider.

  • Supports provider-specific models:
    • Anthropic Claude Models:
      • Claude 3 Series:
        • claude-3-haiku: Fastest response times, ideal for tasks like customer support and content moderation
        • claude-3-sonnet: Balanced performance for general-purpose use
        • claude-3-opus: Advanced model for complex reasoning and high-performance tasks
      • Claude 3.5 Series:
        • claude-3.5-haiku: Enhanced speed and cost-effectiveness
        • claude-3.5-sonnet: Superior performance with computer interaction capabilities
    • OpenAI: gpt-4o, gpt-4o-mini, gpt-4-turbo
    • OpenRouter: Any model in provider/model format (e.g., openai/gpt-4, anthropic/claude-2)
    • Ollama: Any locally available model (e.g., llama2, codellama)
  • Persists model selection.

Development

Install dependencies:

npm install

Build the server:

npm run build

For development with auto-rebuild:

npm run watch

Installation

To use with Claude Desktop, add the server config:

• On MacOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• On Windows: %APPDATA%/Claude/claude_desktop_config.json

Example config:

{
  "mcpServers": {
    "letta-memgpt": {
      "command": "/path/to/memgpt-server/build/index.js",
      "env": {
        "OPENAI_API_KEY": "your-openai-key",
        "ANTHROPIC_API_KEY": "your-anthropic-key",
        "OPENROUTER_API_KEY": "your-openrouter-key"
      }
    }
  }
}

Environment Variables

• OPENAI_API_KEY — Your OpenAI API key
• ANTHROPIC_API_KEY — Your Anthropic API key
• OPENROUTER_API_KEY — Your OpenRouter API key

Debugging

Since MCP servers communicate over stdio, debugging can be challenging. Use the MCP Inspector:

npm run inspector

The Inspector provides a URL to access debugging tools in your browser.

Recent Updates

Claude 3 and 3.5 Series Support (March 2024)

• Added support for latest Claude models:
  • Claude 3 Series (Haiku, Sonnet, Opus)
  • Claude 3.5 Series (Haiku, Sonnet)

Unlimited Memory Retrieval

• Added support for retrieving unlimited conversation history.
• Use { "limit": null } with the get_memory tool to retrieve all stored memories.
• Use { "limit": n } to retrieve the n most recent memories.
• Default limit is 10 if not specified.