Claude Dreaming Explained: Agent Memory, Corrections, and Access

Why memory has to look backward while work moves forward

For Claude to learn from experience, it has to decide what’s worth remembering before it knows what’s next.

Anthropic Engineering published “The components of agent memory” on November 6, 2026, under Agents, Context engineering, and Agent memory. The page opens on a timing problem: an agent has to keep acting, while the value of a detail often shows up only afterward. It quotes Søren Kierkegaard: life can only be understood backward, but it must be lived forward.

Anthropic maps that bind onto human memory. The hippocampus captures experience in the moment so a memory forms quickly and is available right away. Only some of those experiences become long-term knowledge. Much of that consolidation happens offline, during sleep and quiet rest, when the hippocampus replays recent experience and gradually integrates it into the neocortex. Agent memory, the team writes, can use a similar split.

What dreaming does with logs, notes, and the next session

Memory is what allows Claude to learn from experience.

Anthropic uses “agent memory” for knowledge Claude keeps so later work can draw on earlier experience. Agents already capture much of their work in logs or transcripts. Dreaming is the offline pass that reflects on those records: fix mistakes, write new notes, and build representations that were not finished during the session. Anthropic says dreaming has been a useful tool for giving Claude effective memory, including shaping what is loaded before a future session starts.

The page places dreaming next to sleep-time compute, proposed by researchers at Letta and UC Berkeley. That proposal uses idle time between interactions to reorganize existing context. Dreaming takes the same idle-time idea and points it at memory: correct notes against the transcript, organize a wiki, promote notes up the memory hierarchy, focus memory on a specific goal, and fill gaps a session leaves behind.

• Correct what was written

  Dreaming can revise memory after the fact, using the transcript to catch mistakes in notes.

• Build what the session skipped

  It can create notes, new memory representations, and a more organized store such as a wiki.

• Aim memory at a goal

  Anthropic describes dreaming that promotes notes up a hierarchy, narrows memory around a specific goal, and fills gaps the live session left open.

• Prepare the next start

  The same pass can shape what is preemptively loaded, so Claude finds what it needs more easily the next time.

Salience, steering, and what the page leaves open

We’ve found that dreaming has been a useful tool for helping us give Claude effective memory.

The engineering page says Claude has learned to judge salience, and it frames memory as a prediction about what the future will need. Steering covers how Claude is guided to write memory in the first place, so the notes that survive are the ones a later task is likely to use. Dreaming then reorganizes that store once the session is over: corrections, new representations, and a tighter set of notes aimed at what comes next. Anthropic closes on the same forward-and-backward theme it opened with.

You can read the full account on anthropic.com. The page explains the method. It does not name a parameter count, a benchmark score, a price, a rate limit, an open-weights release, or a separate API for dreaming. Those product details stay outside this engineering write-up, so the concrete claim you can check today is the process Anthropic describes, not a public toggle or a billing tier.