I prefer to minimize tribal knowledge of all forms. This does incentivize deletion over "it might be somewhere/in someone's mind", but it also incentivizes moving the knowledge out of the jungle and into something longer lived and searchable in a broader context. How much I prefer the minimization depends on the size/stage of the company, though... At early startups for instance, most tribal knowledge will take care of itself by being forgotten -- and no harm either since the subject matter it applied to is no longer relevant, that was x months ago and nearly everything has changed.
For larger companies, though, it's really nice to be able to pull up documentation (glorious if you can even use a public Google search instead of some intranet lookup or README) about something, instead of having to go on a safari to try and find someone who knows someone who might remember why X, find out they don't remember (or don't remember enough/don't have time for you until Later), and having to figure it out from scratch yourself like you hoped to avoid. Not having to do that for everything is such a nice experience that I'd like to encourage more of it where I can.
Coming to the specifics you highlighted, "there's probably a better way to do this", "maybe look at this edge case", to me those are topics for the code review. They can be rephrased as questions for the reviewer(s): "Do you think this could be done in a better way?", "Should we try and test this edge case or does it matter?" The answers are in the review, and possibly in follow up work items tracked by something.
Code reviews can contain a lot of useful knowledge. If your commits aren't trivially linked to a review, it's worthwhile to spend the extra 10 seconds and manually link them at the end for posterity. If you aren't doing reviews, or the reviews are crap, oh well, you asked and no one answered/cared and that's now visible. Don't leave the question in the code.