Anthropic breaks down the components of agent memory, using sleep-time 'dreaming' to organize it

hugobowne · x · 2026-11-08

Anthropic's engineering team published a deep dive on agent memory, opening with Kierkegaard: life can only be understood backward but must be lived forward — the same bind in designing agent memory. Borrowing the hippocampus–neocortex division of labor, agents capture experience in logs while a consolidation phase distills it into long-term knowledge.

Key points:

Dreaming: inspired by sleep-time compute (Letta / UC Berkeley), Claude uses idle time between interactions to reflect on logs — correcting mistakes, writing notes, building new memory representations, and deciding what to preemptively load into future sessions;
The team walks through several evals they use to measure what each memory component contributes.

One of the most systematic public writeups on agent memory engineering to date.

Original post →

Anthropic breaks down the components of agent memory, using sleep-time 'dreaming' to organize it

More from coding & agent