Activity Frames makes capture compilation deterministic
Its guarded replay and single-user study are different from universal UI rollback or elimination of agent memory growth.
Activity Frames compiles passive screen capture into typed episodes without a model in the compiler. Its measured byte-equality contract excludes the generation timestamp: identical captured input then produces the same compiled content and evidence pointers. This makes the record auditable; it does not freeze a live application’s state.
On one professional’s 128756-frame corpus over 51 active days, the paper reports an 86-fold reduction of a representative day’s capture to context and 68 ms compilation for a lighter day; that timing is not a universal full-history latency. Its QA study uses 64 questions over eight days with one answering pass per model tier, reporting 98.4% accuracy against its oracle versus 66–80% for an LLM summary. The raw and summary baselines cannot fit the busiest day and report on seven days, while the compiled block reports on all eight. Its zero-token replay demonstration requires a guard-matched hit; the near-8% token-savings ceiling is modeled from this single-user corpus, not measured provider-bill savings.
Catastrophic Remembering is a separate study of accumulated instructions, reporting 226% growth and its tested comment-removal intervention. A capture compiler is not itself a solution to every instruction conflict or a guaranteed rollback mechanism.
For a routine-replay trial, distinguish captured evidence from executable actions and test guard failures on changing screens. Report actual eligible recurrence and successful replay, not theoretical elimination of all memory cost.
Editorial correction (2026-10-13 UTC): Separated deterministic compilation, guarded execution and the measured delegation ceiling.