Sleep-time compute
Using the time an agent sits idle to reason over its past context before anyone asks, so later answers need less test-time reasoning (Lin et al., 2025). A word problem discussed yesterday gets solved overnight and answered instantly the next day; four scattered facts about a user's colour preferences get rewritten into one ranked list.
In michael-ryan's lecture (Stanford CS329Z, Oct 2026) it is the main example of memory maintenance. On the AIME math benchmark it needs far fewer test-time tokens than sampling many answers, though that comparison leaves out the sleep-time tokens. Ryan's answer to whether it is more accurate: not if you could spend the same compute on the real question, but under a fixed answer budget it can be. The hard part is guessing what to think about, since the questions are unknown, and a sleep-time agent given tools would raise safety questions about what it does while nobody is watching.
It comes from the same line of work as memgpt. It also fills the gap andrej-karpathy describes in params-vs-context-memory, the missing analogue of "updating parameters while you sleep", though in context instead of in the weights.
Source: report