When Rewind AI shifted focus away from desktop screen indexing toward hardware pendants, it left thousands of power users searching for a reliable, local-first alternative. Here is how we're designing Maple to solve screen recall without the pitfalls.
Continuous video recording was the wrong technical primitive. It bloated hard drives and drained batteries. Semantic OCR text indexing captures meaning in a fraction of the space.
Rewind demonstrated that knowledge workers crave an infallible memory of everything they see. However, continuously encoding video frames resulted in 30GB+ monthly storage footprints and significant thermal throttling.
Our Development Architecture: Event-Driven OCR
In Maple, we don't save video streams. Instead, our daemon captures active-window text snapshots only when meaningful screen changes occur, converting pixels into searchable vector embeddings. A full month of memory takes under 2 GB of compressed SQLite storage.