Sleep-time compute was introduced a while ago in Letta arxiv.org/abs/2504.13171 - the main difference is that sleeptime agents are now orchestrated client-side, with the ability to manage context through bash and git
Stateful agents with real memory is the thesis everyone's converging on. Keeping it model-agnostic is the harder bet. Amiko twins sit one layer up so memory travels between Letta, Claude, OpenAI, whoever's on top.