Honestly, I like the idea behind the feature. Sure, I could (and probably should!) be more proactive with my bookmarks, or otherwise saving content I find interesting.

But the fact is, I don't. If something like this (with negligible enough compute/storage impacts) could be implemented locally, I would probably buy into it. I'm sure it would be "better" with a larger model, but as a jumping-off point for finding something I saw or read about prior, I think a locally available model would suffice.

Trusting Microsoft (or Apple, OpenAI, etc.) with this information for a "better" AI model is not a trade-off I'd personally find acceptable.