Building a Privacy-First On-Device AI Memory Assistant for Android

Short summary

A developer built Aye-Aye, a fully local AI memory OS for Android that passively indexes screen content via accessibility services, notification listeners, and offline OCR. It uses a 37 MB on-device embedding model with SQLCipher FTS5 for hybrid keyword-plus-semantic search via Reciprocal Rank Fusion. The post details the architecture challenges of running RAG entirely on-device and invites beta testers.

• •Aye-Aye indexes screen context locally using Android accessibility services, notifications, and offline OCR
• •Hybrid search combines SQLCipher FTS5 keyword search with a 37 MB bge-small embedding model for semantic queries
• •Zero cloud, zero telemetry; developer is seeking beta testers for the privacy-first app

Generated with AI, which can make mistakes.