Sleep-Based Learning: Continual Memory Consolidation in LLMs

This repository contains the implementation of "Sleep-Based Learning," a bio-inspired architecture that mimics biological REM cycles to achieve continual learning in Large Language Models. https://gallahat.substack.com/p/llm-sleep-based-learning-implementing

Features

• Wake State: Standard chatbot interaction.
• Sleep State: Generates synthetic future trajectories ("dreams") based on recent interactions using a specialized "Dream Generator" model.
• Memory Consolidation: Fine-tunes the model on a mixed batch of new dreams, recalled memories (old dreams), and grounding data to prevent catastrophic forgetting.
• Bio-Inspired: Mimics the "Dream to Predict" hypothesis of biological REM sleep.

Repository Structure

• FullPipeline.ipynb: The main notebook containing the full Sleep-Dream-Wake cycle implementation, including the Gradio interface.
• DreamsGenTrain.ipynb: Notebook for training the "Dream Generator" model.
• grounding-dataset.json: Dataset of general facts used for grounding during the sleep cycle.

Installation (python 3.11.14)

1. Clone the repository:

   git clone https://github.com/Gal-Lahat/sleep-based-learning.git
   cd sleep-based-learning

2. Install the required dependencies:

   pip install -r requirements.txt

3. Set up your Hugging Face token:

   • Create a .env file in the root directory.
   • Add your token: HF_TOKEN=your_huggingface_token_here

Usage

1. Training the Dream Generator (optional)

Open DreamsGenTrain.ipynb to train the specialized model responsible for generating synthetic dreams. This notebook demonstrates how to fine-tune a model (e.g., Gemma-3 4B) to predict future conversation turns.

2. Running the Full Pipeline

Open FullPipeline.ipynb to run the interactive chatbot.

• Run all cells to initialize the model and the Gradio interface.
• Chat with the model.
• Type /sleep to trigger the Sleep-Dream-Wake cycle, where the model will generate dreams, train on them, and consolidate memories.

How it Works

1. Interaction: The user chats with the model.
2. Sleep Trigger: When /sleep is invoked, the current conversation history is processed.
3. Dream Generation: The Dream Generator creates synthetic future scenarios based on the conversation.
4. Mixed-Batch Training: These new dreams are mixed with old dreams (recalled from storage) and grounding facts.
5. Fine-Tuning: A lightweight LoRA adapter is trained on this mixed dataset.
6. Merging: The adapter is merged into the base model (or a continuous learning adapter), updating the model's long-term memory.

Citation

If you use this code or concept, please cite:

Gal Lahat. (2025). LLM Sleep Based Learning: Implementing REM-style Cycles and Synthetic Dreaming for Continual Memory Consolidation in LLMs. https://gallahat.substack.com/p/llm-sleep-based-learning-implementing

License

This project is licensed under the MIT License - see the LICENSE file for details.