ContextD captures screen activity for LLMs

This feels like a personal memory layer for the desktop, which is exactly why it’s compelling and a little unnerving at the same time. The real product question isn’t whether it works, but whether users will trust an always-on context collector.

• –The 2-second screenshot loop plus keyframe/delta OCR is a smart compromise between fidelity and storage overhead.
• –No images are stored, which helps privacy and disk use, but makes OCR quality the whole game.
• –FTS5 search and two-pass enrichment make it more than a log viewer; it becomes a context backend for other tools.
• –The localhost API is the multiplier here, because it lets automations and agents consume the memory without going through the UI.
• –OpenRouter keeps the model layer flexible, but screen-derived text still leaves the machine for summarization.