Soulkeeper
Local-first macOS screen recorder with full-text search over OCR-extracted content.
The Problem
You can't search what you saw on screen yesterday. Existing solutions either upload your screen to the cloud or don't search the content. This is a local-only personal memory layer for your machine.
What I Built
A Python-based macOS daemon using CoreGraphics + Apple's Vision framework for OCR, SQLite + FTS5 for full-text search, FFmpeg with hardware acceleration (VideoToolbox) for H.265 encoding. Frames captured every 2 seconds (configurable), deduplicated via perceptual hashing, OCR'd locally, then bundled into chunked video for compact storage. FastAPI web UI for timeline navigation and a CLI for power users.
Notable
Aggressively dedupes frames before OCR using perceptual hashes; captures only the active window by default to minimize redundancy. Video chunks are automatically extracted and JPEGs deleted after chunk verification, keeping storage under 20GB/month for typical use. Privacy-first by construction: nothing leaves the device. Named after the Culture-series device for continuous mind-state recording.
Stack
Status
Source ready; public repo pending.