What's this about?
The benchmark contains 27,218 messages from ten real people talking with AI companions for up to 120 days, with evidence-backed labels for profiles, chat responses and memory questions.
What exactly is longitudinal memory?
It means recognizing when something said weeks or thousands of messages ago matters now, while ignoring old details when they do not.
Why does this matter?
Only 3.4% of messages needed history, so recalling everything is also a failure. Tested systems struggled to detect relevance and inferred traits beyond what users revealed, exposing both retrieval and privacy problems.