What's this about?
A benchmark testing whether personal agents that retain preferences, task histories, tool routines, and learned skills actually get better over time — the empirical foundation of "recursive self-improvement," which has been more assumed than measured (18↑).
What exactly is "recursive self-improvement" here?
The modest, concrete version: an agent turning accumulated experience into better future behavior — remembering you hate clarifying questions, reusing a workflow that worked. Not the sci-fi runaway kind; the "does my assistant get smarter about *me*" kind.
Why does it matter?
It puts a number on a claim products already make ("it learns you"). If retained experience doesn't measurably improve behavior, a lot of "personalized agent" marketing is unbacked — and it connects directly to yesterday's cross-session personalization paper.