memorA is a fully offline, voice-activated AI companion built to assist Alzheimer’s patients on edge hardware like the Raspberry Pi 5. All processing runs 100% locally after setup — no cloud, no internet, no privacy concerns.
Used Github Copilot for inline completions and writing code autonomously and debugging.
Just finished the initial working build of memorA - a fully offline AI companion for Alzheimer patients on a Raspberry Pi 5.
The core loop: WebRTC VAD listens on the mic, segments speech, pipes audio through Faster-Whisper (base, INT8) for transcription, sends text to Qwen 0.8B via Ollama, and speaks back with Piper TTS. 100% offline, zero cloud, full patient privacy.
First-boot orientation asks 3 questions (name, last meal, day of week) and saves a local profile injected into every LLM system prompt.
Incident detection: 30ms RMS checks catch sudden loud noises (falls, glass breaking, screams). Bilingual distress keywords (EN + RO) trigger emergency dialogue - if unresponsive after 8s, logs and simulates 112 call.
Caretaker web UI (Flask, port 5000): view patient profile + incident logs, trigger re-evaluation, set Bluetooth MAC, toggle EN/RO language. Background thread, never blocks the AI loop.
Systemd service + 3AM auto-updater + Bluetooth autoconnect. install.sh handles everything end-to-end.
Next: RPi 5 hardware testing, VAD/RMS tuning, ESP32 wristband for fall detection.
Log in to leave a comment
amazing project! this has lots of potential, and I’m looking forward to seeing this become an actual product!