An AI agent is software that works toward a goal by taking actions in a digital environment, such as clicking through a website or sending commands to a system. When an agent tackles many tasks in a row, it can save what it learns so it does not rediscover everything each time. The trouble is that environments change without warning, and older notes go stale.
The researchers built a test under a name meaning a scope for reviewing experience, with tasks spread across 18 interactive environments. They separated two kinds of knowledge: how a place is laid out and reached, and the rules that govern what you may do there. They then changed things silently and checked both whether the agent finished its task and whether its saved notes stayed correct.
Across 20 AI models, saving experience improved task success. But agents often kept outdated details while failing to update them. Even with correct notes, agents sometimes did not finish. It matters because long-running assistants must maintain their knowledge, not just accumulate it.