Anthropic reported on November 6, 2026, that offline review of an AI agent’s records can help it correct stored mistakes and organize information for future tasks. The company calls this process “dreaming.” Its internal evaluations found benefits in some settings, alongside cases where elaborate memory structures offered little advantage.
Here, memory means notes retained between sessions and read back as context, rather than changes to a model’s trained parameters. Dreaming involves revisiting transcripts or documents to check notes, repair errors and create linked reference pages. The sleep analogy describes an engineering approach; these experiments do not demonstrate biological sleep in AI.
One evaluation used approximately 350,000 internal files, totaling about 420 million tokens, from Stainless, a company Anthropic said it acquired in May. Opus 5 agents built a linked wiki over successive sessions. Across 25 questions, answers using only the wiki rose from meeting 13% of graded criteria after two construction sessions to 48% after 128 sessions.
With document search also available, scores increased from 54% to 61% as the wiki grew, compared with 53% for search alone. The final combined result was eight percentage points above search alone. Anthropic also reported approximately 20,000 fewer tokens used per question, but did not report total wiki construction and maintenance costs, leaving net savings unresolved.
Benefits depended on the task. In a 24-question merger evaluation, wiki-assisted scores on items requiring synthesis across documents, sometimes with judgment or arithmetic, reached 87%, versus 73% with search alone. The size of that subset was not supplied. One Harvey LAB legal-brief task showed no clear gain, while another produced a wiki larger than its source documents.
In an acquisition-contract task, revised wiki-building instructions increased coverage from 41 of 57 graded facts, or 72%, to 57 of 57, or 100%. The revision combined separate contract pages, one agent per contract and verbatim quotations. Because instructions were changed after examining failures, without a reported held-out test, the improvement may reflect optimization for that particular task.
Other tests examined how memory fails. In a Pokémon Red example, reviewing a transcript corrected a mistaken location note and enabled Claude to reach the next city; runs with the original note or no memory did not. On a CL-Bench database task, each question began in a fresh session with access to persistent notes and correct-answer feedback. Prompting Sonnet 4.6 to read those notes and test hypotheses reportedly nearly doubled its score, although exact scores were absent from the supplied text.
These are company-reported engineering results; the supplied article does not establish peer review or independent replication. It provides no confidence intervals, comprehensive run counts or sufficient grading detail to assess statistical robustness. Successive wiki-building sessions and multiple criteria within questions are dependent observations, not independent replications. Offline organization also consumes compute and can retain or introduce errors. The findings support further testing of memory design, without establishing broad superiority or reliable long-term recall.
Cover illustration generated with AI; conceptual, not research evidence.