AMM - Adaptive Memory Manager

An intelligent memory system that provides continuous learning capabilities for AI conversations.

Core Features

• Automatic Memory Injection: The system automatically retrieves and injects relevant memories without requiring explicit user prompts

• Semantic Search: High‑quality semantic understanding based on Gemini 2.0 Flash embeddings

• Continuous Learning: Learns from every conversation to avoid repeating mistakes

• Verifiability: Tracks memory usage and quantifies system improvements

Related MCP server: Amber

Quick Start

1. Install Dependencies

pip install -r requirements.txt

2. Configure API Key

Create a .env file:

GEMINI_API_KEY=your_api_key_here

3. Start the MCP Server

python src/server.py

4. Configure in Claude Desktop

Edit claude_desktop_config.json (see docs for the location) and add:

{
  "mcpServers": {
    "amm": {
      "command": "python",
      "args": ["C:/Users/notli/Desktop/artificial intelligent/AMM/src/server.py"]
    }
  }
}

Project Structure

AMM/
├── src/
│   ├── server.py           # Main MCP server program
│   ├── memory_store.py     # Memory storage logic
│   ├── embeddings.py       # Gemini embeddings interface
│   └── utils.py            # Utility functions
├── data/
│   └── memories.json       # Memory data storage
├── tests/
│   └── test_basic.py       # Basic tests
├── .env                    # API configuration (not committed to Git)
├── .gitignore
├── requirements.txt
└── README.md

Usage

MCP Tools

1. add_memory - Add a new memory

2. search_memory - Search for relevant memories

3. list_memories - List all memories

4. delete_memory - Delete a memory

5. get_stats - View usage statistics

Automatic Injection Mechanism

On each conversation, the system will automatically:

1. Analyze the semantics of the user message

2. Retrieve the 5 most relevant memories

3. Inject these memories into the AI’s context

4. Extract new memories from the conversation

Roadmap

• Phase 1: Basic MCP server + JSON storage

• Phase 2: Automatic memory extraction and management

• Phase 3: Memory lifecycle management

• Phase 4: Vector database integration

Tech Stack

• Language: Python 3.10+

• MCP: Python MCP SDK

• Embeddings: Gemini 2.0 Flash

• Storage: JSON → SQLite → Vector DB

License

MIT License