Reminisc

Reminisc is an OpenAI inspired open-source memory framework for LLMs.

Features

• Stores user input as memories for future reference
• Retrieves relevant memories based on user queries
• Classifies user input to determine if it should be stored as a memory
• Generates contextual responses using retrieved memories and chat history
• Utilizes an LLM for memory creation and response generation
• Integrates with Supabase and pgvector for efficient memory storage and retrieval

Installation

1. Clone the repository:

git clone https://github.com/yourusername/reminisc.git
cd reminisc

2. Install the required dependencies:

pip install -r requirements.txt

3. Set up environment variables:
   • Create a .env file in the project root directory.
   • Add the following variables to the .env file:

OPENAI_API_KEY=
SUPABASE_URL=
SUPABASE_KEY=

Supabase Setup

1. Initialize Supabase: Ensure you have the Supabase CLI installed. If not, install it by following the instructions here.

2. Apply database migrations: Reset the Supabase database and apply the migrations to set up the required tables and functions:

supabase db reset

This will execute the SQL files in the supabase/migrations directory, setting up the necessary database schema.

Usage

To start Reminisc:

1. Run the FastAPI server:

fastapi dev main.py

2. In a separate terminal, run the Streamlit app:

streamlit run app.py

You can now interact with the assistant by entering your messages in the Streamlit app. The assistant will generate responses based on the stored memories and the current conversation context.

API Documentation

For detailed information about the available API routes and their usage, please refer to the documentation.

Configuration

You can modify the configuration settings in the reminisc/config/config.py file. The available settings include:

• CLASSIFIER_MODEL_NAME: The OpenAI model used for classifying user input (default: "gpt-3.5-turbo").
• MEMORY_CREATOR_MODEL_NAME: The OpenAI model used for creating memories from user input (default: "gpt-3.5-turbo").

Contributing

We welcome contributions to improve Reminisc and expand its capabilities. Please open an issue for suggestions or bug reports, or submit a pull request for code contributions.

Priority areas for contributions:

• Packaging: Convert this framework into a distributable Python package.
• LLM Integrations: Integrate additional language models for generating responses.
• Vector Database Support: Expand support for various vector databases for memory storage and retrieval.
• User-Controlled Memory: Allow users to instruct the assistant to remember or forget information and manage memories.
• Memory Consolidation and Updating: Implement mechanisms for consolidating and updating memories to improve response relevance and coherence.

Citation

If you utilize this repository, please consider citing it with:

@misc{reminisc,
   author = {Advait Paliwal},
   title = {Reminisc: Memory for LLMs},
   year = {2024},
   publisher = {GitHub},
   journal = {GitHub repository},
   howpublished = {\url{https://github.com/advaitpaliwal/reminisc}},
}

License

This project is licensed under the Apache License.