ProactMem organizes an AI agent’s memory around recent events, checked activity summaries and recurring behavior supported by past observations. A new preprint tests the method on chronological streams from activity, home and code settings. The offline evaluation measures opportunities to offer assistance, with user benefit and interruption burden outside its scope.
ProactMem, a method presented in a single arXiv preprint, organizes memory for language-model agents that offer assistance without a fresh request. It separates recent observations from daily activity summaries and recurring behavior. Summary details are checked against the events they describe. Proposed routines are then replayed against earlier observations before they become available to guide a new suggestion.[1]
The authors built ProactiveStream, an offline test with 4,539 events and 729 reference assistance opportunities across 17 separate streams. It covers physical-activity support, home observations and code-repository work. Each person, home or repository has its own history. Decisions use the current event and earlier observations; the current event enters memory after the decision. A shared control step filters repeated proposals.[1]
The team reports an overall F1 of 52.6 per cent, a measure combining detection of opportunities with the relevance of interventions. That is a 21.4 per cent relative gain over the strongest external comparison method. Method and benchmark were developed together. A separate labeling check changed absolute scores across some scopes, and the offline scores do not measure recipients’ benefit, interruption burden or deployment safety.[1]
