Dreaming

What it is

An AI agent forgets the conversation once a session (a single batch of work) ends. So it leaves notes on verified solutions or next steps and rereads them in the next session. This is called agent memory. Unlike records written on the spot during a session, dreaming is a step in which the agent reviews and organizes accumulated logs and notes during spare time between sessions. In a research post, Anthropic explained that dreaming applies 'sleep-time compute', proposed by Letta and UC Berkeley, to memory organization; sleep-time compute is the idea of reorganizing existing context during idle time between interactions.

How it works

While working, an agent marks important events or saves notes with a memory tool. At that moment, however, it has to judge importance without knowing what will be needed later. In the dreaming step, it checks notes against the original conversation logs to fix incorrect content, and links notes scattered across multiple sessions by topic. It can also group related content like a wiki to make it easier to find later. The next session can automatically load only the essential notes and look up details when needed.

Why it matters

The longer an agent works, the more its performance depends on how it builds up and corrects its memory. Anthropic reported that in an evaluation on internal documents, adding a wiki built through dreaming to raw-text search raised the rubric pass rate from 53% to as high as 61%, and saved about 20,000 tokens per question. The research also distinguishes agent memory from other approaches. Changing model weights through training, in-context learning that only happens within a single conversation, and context compaction, which replaces old conversation with summaries, are not counted as agent memory. Dreaming draws attention because it lets an agent accumulate experience without retraining the model.

Things to know

Dreaming is not always beneficial. Anthropic said the wiki's benefit was not clear on tasks where the answer could be found within a single document. Another problem is that incorrect or outdated notes can carry over to the next session. What to verify, how often, and how to cover that cost remain open design questions. As of now (November 2026), Anthropic has only published experimental results in a research post. Also note that the term 'dreaming' is used in a different sense in world model research, so check what it means in each article.

Example

Imagine giving an agent such as Claude a memory folder and assigning it work across multiple sessions. On the first day, while working with a client's database, the agent writes notes such as 'the date column in this table uses a different format' scattered across several files. When the session ends, the dreaming process runs. It checks these notes against the original conversation logs, separates confirmed facts from guesses, and combines them into a single document called 'DB caveats'. The next day, the new session reads only that summary first and starts work right away. It consults the wiki only when detailed evidence is needed.

Related terms

Agent memory AI agent Context engineering Context window Harness RAG

References

• Designing agent memory - what to remember and how to fix wrong memories (GeekNews)
• The Components of Agent Memory (Hacker News)
• Context Compaction for World Models (arxiv.org)
• Anthropic gives Claude agents a ‘dreaming’ mode to tidy up their memories (Crypto Briefing)
• Anthropic: Claude ‘Dreaming’ Took Wiki Score From 13% to 48% (FourWeekMBA)
• Anthropic Explores AI Memory with Dreaming and Salience Techniques (Blockchain News)