Quick Answer
Anthropic published "The components of agent memory" on November 6, 2026, describing how Claude agents use "dreaming," an offline pass that rereads session transcripts to fix wrong notes, build wikis, and choose what to load next time. On one internal test, scores rose from 13 percent to 48 percent of graded criteria over 128 dreaming sessions.
Key takeaways:
- Dreaming is an offline pass between sessions, not a new model. Claude rereads its own logs, checks notes against what actually happened, and rewrites them.
- Anthropic breaks agent memory into four layers: transcripts, memory notes, representations such as wikis, and the small always-loaded context.
- On Anthropic's own CL-Bench database task, mean reward rose for all six Claude models shown, with Opus 5.5 going from 0.655 to 0.838.
- A wiki built over 128 dreaming sessions from about 350,000 internal documents lifted scores on 25 company questions from 13 percent to 48 percent.
- Dreaming is a research preview on Managed Agents, by request, and is not eligible for zero data retention or HIPAA. It is not a general feature for every Claude user yet.
What is dreaming?
Dreaming is Anthropic's name for what Claude does with its idle time between sessions. In an engineering post published November 6, 2026, titled "The components of agent memory," Anthropic describes how a Claude agent rereads its own session transcripts, fixes notes that turned out to be wrong, writes new notes, builds linked wiki pages, and decides which few facts deserve to be loaded into every future session.
The idea borrows from sleep-time compute, a concept proposed by researchers at Letta and UC Berkeley: models use the quiet periods between interactions to reorganize what they already know, much as the human brain consolidates memories during sleep. Anthropic stresses one point clearly. Dreaming changes no model weights. Memory here means experience-derived context, stored as files Claude can read in later sessions.
What are the four layers of agent memory?
Anthropic defines agent memory as knowledge Claude gains from experience and keeps between sessions as context it can read later. The post breaks it into four layers, from raw record to the small set of facts loaded every time.
| Layer | What it is | Example |
|---|---|---|
| Transcripts | The raw record of conversations and activity | Full session logs |
| Memory notes | Short lessons, rules, and plans Claude writes for itself | "Always check the staging database first" |
| Representations | Organized knowledge built from notes and transcripts | Linked wiki pages |
| Always-loaded context | The small set of facts loaded into every session | A short MEMORY.md index |
Dreaming runs between all four layers. It reviews transcripts, checks notes against what actually happened, fixes errors, writes new notes, and promotes the facts that matter into the always-loaded layer. The design rule Anthropic says works best is simple: tell Claude what memory is for, but not how to use it.
What did Anthropic actually measure?
The post reports Anthropic's own evaluations, not independent measurements, and several are single runs. The headline test is a database lookup task from Continual Learning Bench (CL-Bench) 1.0, run on Claude Managed Agents. Each question started in a fresh session. The only carryover was a memory directory that persisted between sessions. After each answer, Claude was told the correct solution and could write a memory. Anthropic then ran dreaming on that memory folder.
| Model | Mean reward before dreaming | Mean reward after dreaming |
|---|---|---|
| Opus 5.5 | 0.655 | 0.838 |
| Sonnet 5.5 | 0.644 | 0.833 |
| Opus 5 | 0.564 | 0.828 |
| Opus 4.7 | 0.383 | 0.700 |
| Sonnet 5 | 0.306 | 0.610 |
| Sonnet 4.6 | 0.330 | 0.364 |
Mean reward rose for all six models shown. The gains are large for most models, and small for Sonnet 4.6, which moved only from 0.330 to 0.364. That uneven result is worth noting. Dreaming helps, but it does not help every model equally.
How did a wiki go from 13% to 48%?
The most striking number in the post comes from a wiki experiment. Anthropic built a wiki from about 350,000 internal Stainless files, roughly 420 million tokens. Over 128 dreaming sessions with Opus 5 agents, performance on 25 evaluation questions grew from 13 percent to 48 percent of graded criteria. Adding search on top of the wiki reached 61 percent, against 53 percent for search alone.
That is a big jump, and the caveats matter as much as the number. The corpus is Anthropic's own internal documents, which outside readers cannot inspect. The evaluation questions are Anthropic's own. The result shows that dreaming can organize a large internal knowledge base into something an agent can actually use, but it is not a benchmark you can reproduce at home.
What is the Pokemon story?
Anthropic makes the failure mode concrete with a game. Claude was playing Pokemon Red and used a Pokemon Center near the entrance to Mount Moon as a landmark. The note was wrong, and the wrong note kept steering later sessions into failure, because Claude trusted its own memory.
The researchers ran dreaming on the folder containing the incorrect note. Claude checked the note against the original session transcript, caught the error, and rewrote the memory. The corrected branch re-entered Mount Moon and eventually reached Cerulean City.
The story is small, but the point is large. Agent memory can carry a mistake into every fresh session. A wrong note is worse than no note, because the agent trusts it. Dreaming exists to catch exactly that kind of error, by checking notes against the transcript they came from.
How do developers use dreaming today?
Dreaming is available as a research preview on Claude Managed Agents, and developers request access. The Managed Agents documentation describes a dream as an asynchronous job that reads an existing memory store and 1 to 100 past sessions and produces a new, reorganized store, without changing the input. It usually takes minutes to a few hours. Calls need the beta headers managed-agents-2026-04-01 and dreaming-2026-04-21.
There are two hard limits to know before you plan around it. Dreams are not eligible for zero data retention or HIPAA, so data with those requirements cannot be used with them for now. And the post describes research setups on Managed Agents and Claude Code's auto memory. It does not announce dreaming as a general product feature, so it is not something every Claude user can turn on today.
Claude Code already has its own memory behavior. Its auto memory loads only the first 200 lines or 25 KB of a MEMORY.md file as an index, whichever comes first. That limit is why the always-loaded layer must stay small, and why dreaming's job of choosing what to promote matters.
What should builders copy from this?
You do not need Anthropic's infrastructure to use the idea. The design lessons in the post work at any scale. Keep memory as files your agent can read, not as hidden state. Let the agent write short notes about what it learned, then run a separate pass that checks those notes against the session transcript and fixes the wrong ones. Keep the always-loaded index tiny, and let the rest live in notes and wikis the agent can look up when it needs them.
The most useful rule is the one Anthropic states plainly. Tell Claude what memory is for, but not how to use it. Give the agent the purpose, and let it decide the mechanics. That is the difference between a memory system that helps and one that fills context with noise.
Sources: Anthropic, The components of agent memory, engineering blog, November 6, 2026; Anthropic Managed Agents documentation, Dreams; Anthropic Claude Code memory documentation; Letta and UC Berkeley, sleep-time compute, April 2025.
Related reading: AI Productivity & Automation, Claude Message Threads Cut Agent Wait by 35 Percent, Anthropic Watermarks 4 More Claude Models
Frequently Asked Questions
What is Claude dreaming?
Dreaming is an offline pass in which Claude rereads its session transcripts between sessions to fix wrong notes, write new ones, build wikis, and choose which facts load into every future session. It changes no model weights.
When did Anthropic publish the agent memory post?
On November 6, 2026, on its engineering blog, under the title "The components of agent memory," written by Lance Martin.
How much did dreaming improve scores?
On Anthropic's own CL-Bench database task, mean reward rose for all six models shown, with Opus 5.5 going from 0.655 to 0.838. A wiki built over 128 dreaming sessions lifted scores on 25 internal questions from 13 percent to 48 percent of graded criteria.
Can I use dreaming in Claude today?
Dreaming is a research preview on Claude Managed Agents, available by request, using the beta headers managed-agents-2026-04-01 and dreaming-2026-04-21. It is not announced as a general feature for every Claude user.
Does dreaming work with zero data retention or HIPAA data?
No. Dreams are not eligible for zero data retention or HIPAA, so data with those requirements cannot be used with them for now.
What is the Pokemon story about?
Claude used a wrong landmark, a Pokemon Center near Mount Moon, as a memory note while playing Pokemon Red, and the error kept steering later sessions wrong. A dreaming pass checked the note against the transcript, corrected it, and the agent eventually reached Cerulean City.


