Anthropic describes Dreaming as an offline process that checks Claude’s experience-based memory against session transcripts, revises or prunes notes, and selects information for later sessions. The approach is meant to maintain memory between interactions; Anthropic’s reported evaluations show that its usefulness varies by task. Anthropic’s engineering post, published November 6, 2026, describes the mechanism and results.

What Dreaming does to agent memory

A three-stage flow links an active Claude session to Dreaming’s memory review, then to selected information entering a later session.

Here, persistent memory means knowledge Claude retains from interactions and assigned tasks for use in later sessions. During a session, Claude can record observations in memory notes. Dreaming works afterward: it can check those notes against the original transcripts, correct or remove entries, and organize information into representations such as linked wiki pages. It can also influence which memories are brought into a later session.

That is different from changing a model’s weights, learning within one conversation, or compressing earlier turns into a summary. Dreaming manages experience-based context; it does not retrain the model.

Anthropic also connects this division of work to sleep-time compute, an approach proposed by researchers at Letta and UC Berkeley for reorganizing context during idle time. Anthropic describes Dreaming as applying a similar idea to agent memory.

A memory system earns its place when it preserves information that stays useful across sessions; adding more notes alone is not a win.

When memory helps—and when it can get in the way

Anthropic reports gains in some tasks that require carrying useful information across sessions or combining material from multiple documents. Its results also show that a wiki is not automatically helpful: other tasks showed no clear gain, or produced a wiki larger than its source documents.

EvaluationSetupAnthropic’s reported result
Stainless corpus, 25 questionsAbout 350,000 files and 420 million tokens; Opus 5 agents built and revised a wiki across successive sessionsWiki-only score rose from 13% after two sessions to 48% after 128 sessions
Stainless corpus, search and wikiThe same corpus, with search alongside a growing wikiCombined score rose from 54% to 61%; search alone scored 53%. Anthropic reported savings of about 20,000 tokens per question with the combined method.
Merger questions24 questions, many requiring information from multiple documents87% with the wiki, compared with 73% with search alone
Contract termsA shared summary page was replaced by one page per contract, with terms quoted verbatimThe shared summary retained 41 of 57 graded facts (72%); the revised wiki retained all 57 (100%)

The pattern depends on what a task demands. A structured memory can help when an agent must connect details scattered across documents. In one Harvey LAB task, where criteria could often be found in a single document, Anthropic reported no clear gain over text search. In another, the wiki grew larger than its source documents and abstracted little.

Memory also has a carrying cost. Anthropic says Dreaming requires compute, while notes brought into a session use context tokens. A mistaken note can recur in later sessions, and a once-accurate note can become stale as circumstances change. So the useful question is not simply how much an agent remembers, but whether the saved information remains relevant to future work.

What Anthropic announced about access to Dreaming

On May 6, 2026, an Anthropic reply described Dreaming for Claude Managed Agents as a Research Preview. The dated reply applies to that announcement.