We can log active processes and services, windows' headers and states, their and mouse's position, integrate it with one's git versions' and browser view's history, history of all file relocations done by select programs or\and by user. If there's an AI assistant like M$ Autopilot, also log every request and output in a text form, log keys, back up settings and configs. If we talk about screenshots, pure text table is as light as a feather and is easier to work with, so this 3sec delay looks like an overkill, even though they'd find a way to compress it. With enough data, it's probably easier to take time and reconstruct an approximate screencap than hoard it.

I imagine dragging your position on a timeline across entire months may be a fun novelty. But I don't see myself having a reason to use it and prefer to lose information over logging so much of it even if it's secure.

I mean sure, if you program in folder A for project A and folder B for project B that is easy and doesn't even require any machine learning but I was thinking of research in the browser or writing documents that are not directly labelled with any project information.