A theoretical paper proposes that three phases of a selectionist sleep account correspond to operations in digital memory for persistent LLM agents. It predicts selectionist consolidation outperforming instructionist approaches as memory stores become large, treating sleep as necessary for continual learning under finite capacity.