AI agents that help you proactively, like offering to remind you about your evening walk, often decide based on just recent events or vague memories. They struggle to know when a pattern is real and when it applies to the current moment, so they either miss chances or interrupt you too often.
The researchers built a memory that keeps recent events, longer activity patterns, and recurring behaviors separately. It checks patterns against original events and replays history to see if the behavior is consistent. Only then does it match the pattern to the present situation. They also created a test set with sports, home, and coding scenarios.
Their system scored 52.6% on an F1 measure, which balances catching opportunities against false alarms. That is a 21.4% improvement over the best existing approach. So a machine that remembers and validates recurring patterns can offer help more accurately, without being a nuisance.