Software developer Ralph Pecayo has developed Aye-Aye, an AI-powered memory OS for Android that functions entirely offline to ensure data privacy. By eliminating cloud-based processing and telemetry, the application serves as a personal knowledge base that indexes a user's digital footprint locally on their device.

To capture screen context without requiring manual input, the architecture leverages Android’s native accessibility services to ingest structured text in real-time, utilizing automated filtering to reduce UI noise. A dedicated background service acts as a notification listener to log alerts into a local timeline. For content where structural text is unavailable, a floating manual screen capture tool triggers an offline OCR (optical character recognition) engine to extract text directly from pixel data.

The system employs a dual-engine approach for search and retrieval. Raw data is processed by SQLCipher FTS5 for encrypted, rapid keyword searching with stemming and fuzzy matching capabilities. For semantic queries, the application utilizes a 37 MB bge-small embedding model, which vectorizes user data to support concept-based searches. These results are unified using Reciprocal Rank Fusion to ensure high accuracy.

The final component is a grounded RAG (retrieval-augmented generation) interface that enables users to query their captured history through a floating chat overlay. The system restricts the LLM to answering exclusively from the user's stored local history, ensuring the model does not hallucinate beyond the device's captured data. Pecayo reports that managing memory and battery efficiency while running ingestion, vectorization, and LLM inference entirely on physical Android hardware remains a significant technical challenge. The developer is currently recruiting beta testers to refine performance across various Android hardware profiles.

Software developer Ralph Pecayo has developed Aye-Aye, an AI-powered memory OS for Android that functions entirely offline to ensure data privacy. By eliminating cloud-based processing and telemetry, the application serves as a personal knowledge base that indexes a user's digital footprint locally on their device.

To capture screen context without requiring manual input, the architecture leverages Android’s native accessibility services to ingest structured text in real-time, utilizing automated filtering to reduce UI noise. A dedicated background service acts as a notification listener to log alerts into a local timeline. For content where structural text is unavailable, a floating manual screen capture tool triggers an offline OCR (optical character recognition) engine to extract text directly from pixel data.

The system employs a dual-engine approach for search and retrieval. Raw data is processed by SQLCipher FTS5 for encrypted, rapid keyword searching with stemming and fuzzy matching capabilities. For semantic queries, the application utilizes a 37 MB bge-small embedding model, which vectorizes user data to support concept-based searches. These results are unified using Reciprocal Rank Fusion to ensure high accuracy.

The final component is a grounded RAG (retrieval-augmented generation) interface that enables users to query their captured history through a floating chat overlay. The system restricts the LLM to answering exclusively from the user's stored local history, ensuring the model does not hallucinate beyond the device's captured data. Pecayo reports that managing memory and battery efficiency while running ingestion, vectorization, and LLM inference entirely on physical Android hardware remains a significant technical challenge. The developer is currently recruiting beta testers to refine performance across various Android hardware profiles.