check out this similar project, it's made for Linux and has LLM feature: https://github.com/apirrone/Memento

I’m surprised not more momentum on this. It would be neat to modify this to support local a LLM and maybe with txtai try to build semantic knowledge graphs on screenshots to auto detect “topics” of your desktop activity.

I wonder if there is a sensible API where you can track everything without storing a video file. Something like seeing Window positions, classes, names, and maybe even GTK/QT cache?

You can look into accessibility APIs, or see how Arbtt does its rule-based metadata tracking: https://arbtt.nomeata.de/doc/users_guide/effective-use.html#...

Similar (activity recording of each window), but not exactly the same (no screenshots):

https://github.com/ActivityWatch/activitywatch

Same, I've been looking for something like this ever since Rewind came out.