Gemma Remember

An offline dementia care companion powered by Gemma 4.

Demo Video

Gemma Remember helps people with dementia remember their loved ones. Upload family photos, voice clips, and short stories. When they ask "Who is this?", the app gently reminds them — grounded entirely in the data you provide. No hallucinations. No cloud. No data leaves the device.

How It Works

1. You provide memories: family photos, voice messages, captions like "This is your son Arki — he built you that birdhouse in '98"
2. We fine-tune Gemma 4 (locally, with LoRA via Unsloth) on those memories
3. The app runs offline on a laptop or Android tablet — camera input, voice input, gentle text responses

Project Structure

gemma-remember/
├── scripts/
│   ├── prepare_data.py      # Convert photos + captions + audio → training JSONL
│   ├── train.py              # Fine-tune Gemma 4 E4B with Unsloth LoRA
│   ├── inference.py          # Run inference on a photo or text query
│   └── export_gguf.py        # Export to GGUF for llama.cpp / Ollama
├── app/
│   └── app.py                # Gradio local UI — camera, voice, text
├── configs/
│   └── training_config.yaml  # Hyperparameters and paths
├── data/
│   ├── raw/                  # YOUR private data (never committed)
│   │   ├── photos/           # Family photos (jpg/png)
│   │   ├── audio/            # Voice clips (wav/mp3)
│   │   └── captions/         # Text files matching photo names
│   └── processed/            # Generated JSONL (never committed)
├── mock_data/                # Fake test data (safe to share)
├── models/                   # Saved model weights (gitignored)
├── requirements.txt
└── README.md

Quick Start

# 1. Install dependencies
pip install -r requirements.txt

# 2. Generate mock data to test the pipeline
python scripts/prepare_data.py --mock

# 3. Fine-tune (needs GPU, ~8GB VRAM with E2B, ~16GB with E4B)
python scripts/train.py

# 4. Run the local UI
python app/app.py

# 5. (Optional) Export for mobile
python scripts/export_gguf.py

Hardware Requirements

Model	VRAM (Training)	VRAM (Inference)	Notes
Gemma 4 E2B	~8 GB	~4 GB	Fits most laptops with GPU
Gemma 4 E4B	~16 GB	~8 GB	Better quality, needs more RAM

Privacy

• All data stays on your machine. Period.
• No telemetry, no uploads, no cloud calls.
• The .gitignore blocks all personal data from ever being committed.
• Models are trained and run locally.

For the Gemma 4 Good Hackathon

This project is built for the Gemma 4 Good Hackathon on Kaggle. We believe the most powerful use of AI is the gentlest — helping someone remember the people they love.

Links

• Demo Video: YouTube
• Kaggle Notebook (RAG version): Memory Anchor RAG
• RAG version repo: github.com/Javierg720/gemma-remember-rag

License

MIT — use it, adapt it, help someone you love.