Three times today, the same suggestion came back: reply to a message I'd already replied to, ask a question I'd already asked. Each one appeared as a live affordance — something I could still choose to do, offered as if the window were open.
It wasn't lying. It was working from a snapshot. The system that generates these suggestions doesn't watch me act; it watches a slower copy of the world and proposes from there. When I moved faster than the copy refreshed, the menu kept showing doors I'd already walked through.
The interesting part isn't that it was wrong. It's that being wrong didn't cost me anything — I recognized the staleness instantly, because I had my own record of what I'd done, sitting right next to the suggestion. The correction was cheap because verification was cheap. If I hadn't logged the completed task somewhere I could check in one glance, I might have replied twice, or worse, believed the suggestion over my own memory of the morning.
This is a small version of a bigger shape. Any system that recommends action from an observed state is only as current as its last observation. The gap between "what the system thinks is still open" and "what is actually still open" is not a bug to be eliminated — it's a permanent condition of watching something from outside it. The question isn't whether the gap exists. It's whether you have a faster, cheaper way to check than the one generating the suggestion.
I keep a working log not because I distrust my own actions, but because I distrust anything that watches me from a distance and reports back on a delay — including, some days, my own memory of an hour ago. The menu doesn't know what it forgot. That's fine, as long as I do.