Activity Frames: Deterministic Screen-Activity Compilation for Agent Memory and Replay
A zero-model compiler turns captured screen activity into deterministic agent memory for replay.
The paper says Activity Frames segment local screen captures into typed episodes with app, site, timing, input volume, and evidence pointers. On a single-user corpus of 128,756 frames across 51 active days, it compressed a day of raw capture into a prompt-ready block 86x smaller in 68 ms. An agent using that block answered day-specific questions at 98.4% accuracy, above LLM summaries of the same capture. The compiler also produced measured estimates for routine overhead and recurrence, and replayed a matched routine with zero model tokens.
HF Daily Papers' note
The paper says Activity Frames segment local screen captures into typed episodes with app, site, timing, input volume, and evidence pointers. On a single-user corpus of 128,756 frames across 51 active days, it compressed a day of raw capture into a prompt-ready block 86x smaller in 68 ms. An agent using that block answered day-specific questions at 98.4% accuracy, above LLM summaries of the same capture. The compiler also produced measured estimates for routine overhead and recurrence, and replayed a matched routine with zero model tokens.
HF Daily Papers' note
score 4