New AI Memory Trick Reminds You What You Promised To Do

Right now, when an AI assistant searches your notes, emails, or chat history, it works like a search engine: you type something, and it hands back whatever looks most similar. That's useful, but it's backward-looking. It has no idea you promised to send your accountant a receipt, or that you told a friend you'd follow up next week. So the reminders you actually need never show up — unless you happen to ask for them.

Jonathan Groff's new paper proposes a simple fix: keep a separate, plain list of your commitments, each with a date or a trigger attached ("when the invoice arrives," "on Sunday"). When one of those entries comes due, any related memory gets a small relevance boost. The clever part is that this boost is applied at the last second, blended in with normal search, so it never buries a better answer. And because the list is prepared in advance, it costs essentially nothing to run when you ask a question — no extra computing bill.

In tests using 175 practice tasks built to resemble a public benchmark called TriggerBench, the reminders went from catching zero of the hardest cases to catching nearly all of them, and up to 100% with a stricter setting. Just as important, it raised no false alarms across 53 tasks where commitments were already handled. Groff also found that only about one in five naturally worded promise-trigger pairs actually confuse a normal search system, meaning this fix helps in a real but modest slice of cases.

The honest caveats: the test set was written by the author himself, not an independent group, so treat the numbers as promising rather than proven. Full testing on the real TriggerBench is promised once that data is public. The work is early, but it points at something you'll likely feel in a year or two — an assistant that taps you on the shoulder instead of waiting to be asked.