Memory Twin AI

Memory Twin AI is a consent-based digital memory simulation designed for the AMD Developer Hackathon 2026. Instead of fine-tuning or training massive models from scratch, the system introduces an ultra-fast, explainable Retrieval-Augmented Generation (RAG) architecture running entirely on local AMD compute. By combining Qwen3-Embedding-0.6B and ChromaDB with Qwen2.5-7B-Instruct, the app achieves sub-second memory retrieval and response generation, pulling contextually relevant fictional memories out of a secure JSON vault based on cosine similarity. What makes Memory Twin AI truly unique is its intersection of radical engineering transparency and a warm, human-centric user interface. Every response inside the WhatsApp-inspired chat interface features a completely transparent "Explainable RAG Panel," mapping out the exact retrieval times, confidence scores, and models utilized. This pipeline feeds directly into a reactive 8-bit anime companion panel. Operating entirely via local browser TTS and lightweight Three.js bone rigs rather than slow, resource-heavy real-time video generation, the companion shifts expression, mood, and speech pacing dynamically based on the emotional category of the retrieved memory (such as humor, career, or childhood nostalgia). Built as a complete end-to-end prototype, the system includes a "Memory Atlas" timeline scrapbook, an automated multi-step Hackathon Demo Mode for video presentation, and a dedicated AMD Compute Proof screen that pulls live system metrics directly from the FastAPI backend to prove native ROCm/PyTorch acceleration.