Anthropic details "dreaming" for Claude agent memory
Anthropic has published research on how Claude agents store and consolidate memory between sessions, arguing that offline "dreaming" over past logs measurably improves recall and task performance.
Anthropic published a research post setting out how it thinks about agent memory, meaning knowledge Claude gains from experience and keeps between sessions as context it can read later. The definition leaves out weight updates, in-context learning and compaction.
The post borrows from neuroscience. Agents capture their work in logs and transcripts as they go, much as the hippocampus records experience. A separate offline step, which Anthropic calls "dreaming," then reviews those logs to fix mistakes and write new notes. The idea builds on sleep-time compute, which Letta and UC Berkeley researchers proposed.
On Continual Learning Bench (CL-Bench) 1.0, run through Claude Managed Agents, each question started in a fresh session that could read only a persistent memory directory. Claude Sonnet 4.6 wrote untested guesses spread across several files. Claude Opus 4.7 combined its notes into one document and marked which facts it had verified. Anthropic says Claude Opus 5 was the first model to "close the loop," writing rules along with the evidence behind them and warnings about traps. Opus 5.5 and Sonnet 5.5 did the same and scored about as well as Opus 5 with memory, though both started from slightly higher baselines without it.
In a Pokémon Red test, Claude Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge by roughly step 10,000. The excerpt cuts off before the comparison with Claude Fable 5.
According to the post's summary, dreaming also raised recall on a Stainless wiki task from 13% to 48% and coverage on a legal brief from 72% to 100%. On merger QA, it scored 87% against 73% for search alone. The excerpt doesn't describe how those evaluations were set up.