This paper studies how AI agents get better while they are working, not just whether they finish the job.
The paper tackles understanding super long, first‑person videos (days to a week) by giving an AI a smarter memory and better tools.