Compiling screen capture into agent memory deterministically beats LLM summarization by 18 to 32 points.

arXiv 2608.05784 puts no model in the loop: it compiles passively captured screen activity into a prompt-ready block that's byte-identical, cacheable and mechanically auditable. On 128,756 frames across 51 active days from one professional, it compresses a day 86x in 68ms, and an agent answering questions about that activity reaches 98.4% accuracy (Wilson 95% CI 91.7 to 99.7%) against 66 to 80% for LLM summaries. It also reports Routine Overhead Ratios of 60 to 343x and a delegable-recurrence rate of 9.0% in-sample.