What should an AI memory do when your notes change?

This is the part of Locul we spent the most time on.

Locul reads your notes, docs and dictations and turns them into memories your AI can search. The problem is that notes change. You change your pricing, but the old price is still written in a dozen older notes. If the memory keeps the old one, Claude tells you the wrong price and sounds sure about it.

So when a new note comes in, Locul compares it against what it already knew and decides if an old memory is out of date. Our first version of that check retired too much. We tested it on 151 tricky pairs of similar memories from a real Locul brain, and it retired 26 old memories. 22 of those were still true.

Now it only retires an old memory when the new one contradicts it on the same point and nothing else is lost, or when the new one already covers all of it. If it isn't sure, it keeps both. On the same 151 pairs, that brought the wrong ones down to 4. And when a memory does get replaced, the old version stays. You can turn on "show all" in the Memories tab and see what it replaced.

Once a week it also goes through and removes near-identical duplicates, and merges people or companies that got saved twice.

4 out of 151 still isn't zero, so I'd like to hear how you'd want this to work. Should the newer note always win? Or should Locul ask you when two notes disagree, even if that means more questions?