Dynamic Sleep Schedules for AI Agents

This concept paper proposes the term dynamic sleep schedule for AI agents. Building on recent work on agent memory and sleep-based continual learning, it argues that an agent's inactive periods should be used not only to consolidate memories, but to revisit past decisions across time and learn how much computation a given task deserves. The paper outlines a tiered memory "ladder", a method for evaluating decisions from interaction metadata, selective triggers for review that can also operate while the user is active, and a way to distinguish one-off changes in user behaviour from lasting preferences. The argument is conceptual, developed through an analogy with human exam preparation, and is offered as a hypothesis for further research.