Vector Memory

Vector Memory is a terminal-based AI assistant that uses Google's Gemini for conversational intelligence and Pinecone for persistent, queryable long-term memory.

Features

• Natural Conversation: Engage in natural, human-like conversations.
• Persistent Memory: Remembers key facts from conversations.
• Context-Aware Responses: Retrieves relevant memories to provide contextually-aware, personalized responses.
• Continuous Learning: Continuously learns and appends new memories over time.
• Vector Normalization: Ensures that vector embeddings are normalized for accurate cosine similarity calculations, making the system more robust.

How It Works

The application follows a request/response/memory cycle:

1. User Input: The user enters a message.
2. Memory Retrieval: The application generates a vector embedding of the user's input and queries the Pinecone index to find the most similar memories.
3. Context Assembly: A new prompt is constructed for the chat model, including a system prompt, the retrieved memories, and the user's current message.
4. LLM Interaction: The assembled prompt is sent to the Gemini model.
5. Response Generation: The application receives and displays the generated response.
6. Memory Extraction & Storage: The conversation turn is sent to Gemini in a separate call to extract new, permanent facts. These facts are then embedded and upserted into the Pinecone index.

graph TD
    A[User Input] --> B{Generate Embedding};
    B --> C[Query Pinecone];
    C --> D[Retrieve Memories];
    D --> E{Assemble Prompt};
    A --> E;
    E --> F[Send to Gemini];
    F --> G[Display Response];
    F --> H{Extract New Facts};
    H --> I[Embed New Facts];
    I --> J[Upsert to Pinecone];

Loading

Technologies Used

• Language: Python
• LLM: Google Gemini
• Vector Database: Pinecone
• Libraries:
  • google-generativeai
  • pinecone-client
  • python-dotenv
  • rich
  • numpy

Setup and Installation

1. Clone the repository:

   git clone https://github.com/your-username/vector-memory.git
   cd vector-memory

2. Create a virtual environment:

   python -m venv .venv
   source .venv/bin/activate  # On Windows, use `.venv\Scripts\activate`

3. Install the required packages:

   pip install -r requirements.txt

4. Create a .env file in the root of the project and add your API keys:

   GEMINI_API_KEY="YOUR_GEMINI_API_KEY"
   PINECONE_API_KEY="YOUR_PINECONE_API_KEY"
   

Usage

To run the application, execute the following command:

python vector_memory.py

Customization

• System Prompt: You can customize the AI's personality and behavior by modifying the SYSTEM_PROMPT variable in vector_memory.py.
• Initial Data: You can add your own initial facts to the INITIAL_FACTS list in vector_memory.py to give the AI some starting knowledge.