loop engineering only holds up if the loop can remember. here's the part that decides it: -> the context window resets to a fixed set of files every iteration, so the in-window memory is disposable -> what has to survive lives outside the window: lean anchor files (VISION, PROMPT, MEMORY, SKILL.md) reset each pass, and a durable external store that persists across all of them -> the loop recalls from the store before each pass and writes the outcome after, so run 47 reads what runs 1 through 46 learned -> skip the store and it drifts, repeats finished work, and re-ingests its own early mistakes as truth. context rot hits around 20-30 turns and a loop runs thousands the loop gives an agent persistence across time, memory is what keeps it from forgetting