The key distinction: scalable memory ≠ scalable reasoning. A model can store evicted context in fixed-size fast weights and still fail if it has not spent enough computation transforming that context into a useful state. That is why the “sleep” phase is interesting: it moves