| license: mit | |
| tags: | |
| - mneme | |
| - memory | |
| - weight-injection | |
| - qwen | |
| # Mneme: Neural Episodic Weight Injection Encoder | |
| Trained encoder for the Mneme memory system - injects facts directly into LLM weights. | |
| ## Usage | |
| ```bash | |
| # Clone the repo | |
| git clone https://github.com/Yusuffarhan13/Mneme-v1-mvp.git | |
| cd Mneme-v1-mvp | |
| # Download the encoder | |
| pip install huggingface_hub | |
| python -c "from huggingface_hub import hf_hub_download; hf_hub_download(repo_id='yusuffarhan/qwen-memory', filename='best_encoder.pt', local_dir='mneme_trained')" | |
| # Run | |
| python qwen.py --encoder mneme_trained/best_encoder.pt | |
| ``` | |
| ## Training Config | |
| - **Delta rank**: 16 | |
| - **Target layers**: [4, 8, 12, 16, 20, 24] | |
| - **Encoder**: 768 hidden, 4 layers | |
| - **Base model**: Qwen/Qwen3-4B | |
| ## What This Does | |
| Injects facts directly INTO model weights (no RAG, no prompt injection): | |
| ``` | |
| /remember My name is Yusuf | |
| /remember I work at Google | |
| What is my name? → "Your name is Yusuf" | |
| Where do I work? → "You work at Google" | |
| ``` | |