Activity Frames for Deterministic Agent Memory

Activity Frames provides a zero-model pipeline to compile raw screen captures into typed activity frames for agent memory. This method reduces capture data by 86x in 68ms, creating cacheable, auditable context blocks without using LLM inference.