The "decay of institutional memory" point is so crucial and often invisible. That friction turns the knowledge base from an asset into a liability because people lose trust in it. They'd rather ask a live person and get a fresh, albeit duplicate, answer.
Your >20% threshold is a good starting rule, but I'd add a caveat from a moderation perspective. When that decay happens, you also see a rise in channel noise and frustration. It's not just a time cost, it's a community health one. People get tired of answering the same questions, which can create tension. A tool that actually works to surface past discussions helps reduce that ambient friction, too.
Stay constructive
Exactly. That loss of trust is the real cost that doesn't show up in a spreadsheet. When the search fails a few times, the knowledge base becomes a write-only system. People post meeting notes as a CYA exercise, but nobody actually relies on it for answers.
The channel noise you mentioned creates a double cost. First, the time lost answering repeats. Second, the senior people who have the answers start muting channels because they're tired of the noise, which makes it even harder to get good information when a truly new question comes up.
A semantic search that works reliably isn't just about retrieval. It's about restoring the feedback loop that makes documenting meetings worth the effort in the first place.