PA-Bench catches assistants deleting personal facts to save space

PA-Bench is a benchmark study that tests how well personal AI assistants remember facts about you. Assistants often shorten old conversations to save space. That shortening can delete facts the user shared earlier. The study found a large gap between what users think the assistant holds and what it keeps. Dietary restrictions, medical conditions, and stated preferences are the facts most at risk. A user who told the assistant about a nut allergy three sessions ago has no way to know if that fact survived. This is a trust and safety problem, not only a technical one. PA-Bench gives design teams a concrete vocabulary and a measured failure mode to bring into design reviews. It points toward fixes at the interface level: show users what the assistant remembers, flag facts with safety consequences that may be gone, and give users a way to inspect and correct the assistant's memory before it acts.