Meta's ProPerMem benchmark finds proactive assistants fail at deciding when to interrupt

Meta researchers introduce ProPerMem, a benchmark for memory-driven proactive assistants built from egocentric video and year-long personal memories averaging 618K tokens.

Every baseline scored at most 15%, even frontier models with retrieval-augmented memory.

The paper reports that GPT-5.3-Chat reached 96% when given oracle evidence in the window, but fell to 38% when choosing its own timing and to 15% when using retrieval. The authors say deciding when to interrupt, not just what to say, is where these assistants fail. Meta's ProPerAgent system, trained with reinforcement learning, reached 42.4% accuracy.