What happened
Researchers developed a way for AI agents to build and maintain structured memory of software environments (screens, actions, workflows) instead of starting fresh each time, which makes them roughly twice as successful at completing multi-step tasks. This matters because AI systems that can remember what they've learned about an interface—like a human learning their way around new software—become reliable enough to actually automate real work instead of failing randomly.
Why it matters
This is the first demonstration that AI agents performing long, complex tasks (like navigating websites or software) can actually improve dramatically by remembering what they've learned, rather than treating each session as isolated. That's the difference between an assistant that needs to be re-taught how to use the same interface every time and one that builds persistent knowledge.