Anthropic finds offline memory reviews can help Claude correct stored mistakes

Claude can carry a mistake from one task into the next by saving it as a memory. In research published November 6, Anthropic describes a way to interrupt that cycle: let the model review earlier interactions between sessions, check its notes and reorganize what it keeps. The company calls this offline maintenance “dreaming.”

The term describes a process, not human sleep. Anthropic defines agent memory as knowledge saved between sessions in files or other context Claude can read later. That is different from changing the model’s trained internal parameters, learning within one conversation or shortening a long conversation into a summary.

During a task, Claude can write down information it judges worth keeping. Later, dreaming lets it compare those notes with the original session logs. It can correct errors, create new notes or build linked collections of knowledge. The original record provides something to check against, rather than asking Claude simply to trust its own earlier summary.

A Pokémon Red experiment made the failure concrete. Claude incorrectly noted that a Pokémon Center was east of Mt. Moon, when it was near the western entrance. That mistake led it to believe it had already crossed the mountain. It then walked over a one-way ledge, returning to an earlier route.

Anthropic ran dreaming on the faulty memory folder. Claude checked the transcript, corrected the location and subsequently re-entered the cave, eventually reaching Cerulean City. In the run using the uncorrected note, it fell back onto the earlier route three times and never reached that destination. The intervention repaired a specific stored error; it was not a general test of real-world reliability.

That comparison used one run per model, each about 10,000 gameplay steps in. It compared different models, not just different memory settings. Still, Anthropic observed a practical problem with Sonnet 3.7’s archive: the game kept changing, leaving detailed records out of date. At one point, it spent 17 steps creating and deleting memory files without taking any action in the game.

Fable 5 instead preserved information the screen could not supply: strategies, plans and rules it had worked out. Its central file held the current goal and a plan for tackling gyms. Recording money or location was often redundant because Claude could read those details from the screen on each turn.

Rigid instructions were not always better. On a database task in the CL-Bench 1.0 evaluation, telling Opus 5 and Opus 4.7 to record only failures removed much of memory’s performance benefit. Older Sonnet 4.6, however, benefited from explicit instructions to read its notes and test its explanations for mistakes. Anthropic reports that this guidance nearly doubled its score.

Dreaming can also build a wiki: linked pages that help Claude navigate a growing knowledge collection. Interesting Engineering’s account of Anthropic’s experiments describes a wiki developed across 128 sessions from roughly 350,000 Stainless documents. Performance on 25 evaluation questions rose from 13% to 48%. That remains below half, despite the substantial improvement.

The same account reports 87% accuracy with a wiki on merger questions requiring information from multiple documents, versus 73% with direct document search. Maintenance also consumes computing resources, and excessive information can overwhelm later sessions. The research therefore treats memory as something to select, check and organize—not simply a larger store to fill.