A reason-and-act agent has one input: the whole interaction history, re-rendered and re-sent at every step, so the file it printed at step 4 is still in step 40’s prompt. This is a rendering layer in front of that loop which compresses only what the model sees and leaves the real history untouched — content an earlier turn already displayed becomes a stub pointing at the turn still holding it, and turns the agent itself marks finished fold into a single progress note. Nothing is predicted and nothing is permanently lost: a wrong fold costs one restore. The rendered prefix is rewritten every three steps rather than every step, which holds the prefix-cache hit rate at 87% against the 51% of rewriting continuously, and across five long-horizon benchmarks it cut tokens by 18–53% and cache memory by 39–58% with task resolution unchanged.