The paper introduces Activity Frames, a deterministic, zero-model pipeline that compiles passively captured screen activity into structured memory for computer-use agents. It segments local screen capture streams into typed activity frames—bounded episodes containing metadata such as application, site, timing, input volume, and evidence pointers to raw data—without relying on any machine learning model. This approach addresses the limitation of current agent memory, which records user instructions but not actual user actions, enabling more efficient replay and reducing redundant inference. Read original