Anthropic published an engineering blog post explaining "dreaming," a process that lets Claude tidy up its memory between conversations. Memory here means what the model keeps from experience: notes it writes during a session and reads back later, separate from its training weights. The problem is that a note can be wrong from the moment it is written, or quietly drift as the world changes, and the next session will trust it anyway. Dreaming gives Claude idle time to go back over those notes and check them against what actually happened.
The clearest demo involved a long run of Pokemon. Claude wrote down the wrong location for a PokeCenter near Mt Moon, then kept walking over a one-way ledge and looping back to an earlier route. Left alone, it never finished the journey. With dreaming, it compared the faulty note against the original session transcript, found the mistake, rewrote the memory file, and eventually reached Cerulean City. Anthropic also had Claude build "wiki" pages from a set of documents, and its score on 25 evaluation questions rose from 13 percent to 48 percent across 128 sessions.
What it means for builders: agent memory is turning into a system you maintain, not a log you append. Dreaming can push detail down into a searchable layer and pre-load what is likely to matter next, so a long-running assistant needs less re-explaining every time you open it. It is not free. Reviewing memory burns compute, and Anthropic warns that continuously checking everything is probably impractical, so it suggests focusing dreaming on facts likely to change and tuning how often it runs.