Proactive AI assistants reportedly struggle to know when to speak

A post describing a Meta paper says ProPerMem tests whether AI assistants can interrupt at the right time with information grounded in a year of personal memory. It pairs video with memories averaging 618K tokens and human-verified trigger windows. The post says every baseline scored at most 15%; GPT-5.3-Chat reached 96% with oracle evidence in the window, dropping to 38% when choosing its own timing and 15% with retrieval.