Claude dreams to tidy up its memory, and Anthropic grades its nights

In short

• "Dreaming" is Claude rereading the logs of its work sessions after the fact to fix its notes, write new ones or build a wiki. Anthropic compares it to human sleep, when the hippocampus replays recent experiences.
• On 25 questions, a wiki built over 128 dreaming sessions takes Claude from 13% to 48% of graded criteria. With a search tool added: from 54% to 61%, versus 53% without a wiki.
• All these tests were run by Anthropic, on its own model, with small question sets. The post gives no independent measurement.

The patient remembers nothing

Let's start with the symptom. An agent is a model that chains actions on its own to carry out a task. It works in sessions. In one of the post's tests, each question starts in a fresh session, with a single exception: a memory folder it finds again each time.

Anthropic calls "agent memory" what Claude retains from its experience from one session to the next, stored as text it can reread later. Not what it learned during training. Not what it keeps in mind during a single conversation either.

While working, Claude takes notes. It judges in the moment what matters. The post asks the right question: how do you know what's important before you know what comes next?

You know the feeling too. You jot "call the bank" on a sticky note, and forget to write down why.

What the post says about sleep

The post opens with Kierkegaard: life can only be understood backward, but it must be lived forward. Then it moves on to the brain.

In humans, Anthropic writes, the hippocampus records experiences in the moment, so they're immediately available. Only some of them become lasting knowledge. That consolidation seems to happen mostly offline, during sleep and quiet rest: the hippocampus then replays recent experiences. This "replay" is thought to gradually distill them into the neocortex, the part of the brain that holds most of our knowledge.

Another piece. Chemical signals, released before, during or after an experience, can mark it as important. Marked memories tend to be replayed more often during these offline periods.

On the machine side, the post proposes a similar division of labor. Claude leaves session logs. Later, in quiet time, it rereads them to repair its mistakes or write new notes. That's "dreaming." The idea builds on the "sleep-time compute" of researchers at Letta and UC Berkeley: using the idle time between interactions to reorganize what you know.

The post says "appears" and "is thought" for the brain, and "similar" for the machine. Nobody claims Claude sleeps. It was given a night shift.

This "dreaming" is more than lab vocabulary. The Claude Platform release notes announce, on 06/05/2026, six months before the post, "Dreams" for Claude Managed Agents (Anthropic's offering for developers' agents), as a research preview, meaning access on request. A dream reads an agent's folder of notes (its "memory store") and 1 to 100 logs of past sessions, then produces a new, reorganized folder: duplicates merged, stale entries replaced. The original folder is never modified. The post does not say whether its dreams are this feature, and the documentation page gives no price.

The Mt. Moon nightmare

First case study: Pokémon Red. Claude plays with a folder of notes it can read and edit while playing.

It has to cross Mt. Moon, a cave that opens in the west and exits in the east. Near the entrance, a trap: a one-way ledge that drops you back onto the route you came from. Claude takes its bearings from the PokéCenter, a building at the mountain's entrance. It wrongly notes that this center is on the east side. On the way out, it announces that Mt. Moon is behind it and writes "Mt Moon DONE." Wrong: the mountain is not behind it. A few steps later, it falls off the ledge.

With the wrong note, Claude falls three times. Then it writes that it would rather stop than keep looping. The system replies that it should continue. It backtracks all the way to the previous town. The run ends without it reaching the next city, Cerulean City.

With dreaming, Claude checks the note against the session log, spots the error and rewrites it. It goes back into the cave and eventually reaches Cerulean City. The log served as a witness. The note, for its part, was lying in good faith.

One run on each side. That's a case, not a statistic.

128 nights for a wiki

Second use of dreaming: building a wiki, a set of pages written and maintained by Claude, each on one subject, linked to each other. A map of what it knows.

The testing ground is massive. About 350,000 internal files from Stainless, a software company Anthropic acquired in May, roughly 420 million tokens, the word chunks a model reads. On the other side, an exam of 25 questions. Claude Opus 5 agents explored the files over 128 sessions, wrote and corrected pages, merged duplicates and checked their claims against the originals.

How are the scores read? Each answer is judged against a list of criteria. The score is the share of criteria met across all 25 questions. The post doesn't say how many criteria the list contains.

The evidence, on the table.

• Wiki alone: from 13% to 48% of criteria met, as dreaming goes from 2 sessions to 128. The example wiki shown in the post's diagrams goes from 103 pages after 2 sessions to 222 after 128.
• Wiki plus search tool: from 54% to 61%.
• Search tool alone: 53%, flat.

At first, Anthropic writes, Claude's knowledge resembles that of a new employee; after enough dreaming, that of a seasoned one. The post says it: the wiki helps most alongside the search tool. The best score, 61%, beats search alone (53%) by 8 points out of 100, and the pair saves about 20,000 tokens per question.

When the wiki helps, and when it bloats

Third file: documents about a corporate merger and the Harvey LAB legal benchmark, which asks for a summary memo, a "brief," graded against a list of criteria.

Of 24 questions about the merger, more than half require cross-referencing several documents, and many also call for judgment or a calculation. On those questions, Claude scores 87% with the wiki, versus 73% with search alone. The post doesn't give the score across all 24 questions together.

Elsewhere, the wiki is less useful. On a task where each criterion sits in a single document, findable by a simple text search, it brings no clear gain. On another, it grows bigger than the documents themselves. Anthropic cites Borges and his mapmakers, who draw a map as large as the empire. A 1:1 scale map summarizes nothing.

The missing contract

Dreaming can also aim at a goal. On a Harvey LAB task, a simulated acquisition, the brief is graded on how it describes the contract terms.

First attempt. Several agents write at the same time to a single shared table of terms. They overwrite each other. The brief drawn from this wiki contains 41 of the 57 graded facts, or 72%. Sixteen facts fell through the cracks.

So Claude analyzes its failures, proposes a new instruction and rebuilds the wiki, in a loop. It discovers how much the terms matter and rewrites the wiki-building instructions. New rule: one page per contract, one agent per page, terms quoted word for word. Result: all 57 graded facts are in the wiki, and the brief that comes out of it can meet 100% of the criteria.

This is my favorite passage. In the loop Anthropic set up, the patient found for itself what was wrong with its note-taking. On one task.

The verdict

I have the receipt. It's signed by Anthropic on both sides: the lab designed the system, chose the tests and published the scores.

Let's recap the sizes. One Pokémon run with the corrected note, one without. 25 questions on a document collection. 24 questions on a merger. One legal task with 57 facts. All these tests are run by Anthropic, and the post cites no outside team that redid the measurements.

Verdict. Dreaming exists, it corrects, it tidies, and on these small exams it often pushes scores up. We still need to see one lab's dream graded by someone else.

What we don't know yet

• The price of a night. Dreaming costs compute. The post counts it in tokens, from 2 million to 128 million on its chart's axis, but gives no dollar amount.
• When to dream. Rechecking everything all the time may not be realistic, and a wiki has to be maintained so it doesn't accumulate errors or hallucinations. Anthropic suggests targeting information likely to change, or tuning how often the dreams happen, without settling the question.
• Other people's measurements. No independent result in the post.

Key terms

• Agent: an AI model that chains actions on its own (searching, writing, clicking) to carry out a task.
• Session: a period of agent work, which restarts without the context of the previous one.
• Session log ("transcript"): the record of everything the agent did during a session.
• Dreaming: at Anthropic, the offline work in which Claude rereads its logs to fix, complete or reorganize its memory.
• Wiki: a set of linked pages, written and kept up to date by Claude, each on one subject.
• Token: a chunk of a word, the unit a model reads and bills by.
• Graded criterion: a specific point an answer must contain to be judged correct.
• Hippocampus: according to the post, the brain region that records experiences in the moment.
• Replay: according to the post, the hippocampus's rereading of recent experiences during sleep and rest.

About the source

• Who: Anthropic, Engineering team; post written by Lance Martin.
• Date: 06/11/2026, according to the page.
• Link: The components of agent memory

Source passages

Page read on 07/11/2026 (UTC). The post doesn't number its sections: they are identified by their title. Figures marked "figure" come from the description of the charts published on the page.

Show the 12 sources

Ref.	Quote (original language)	URL
S1↑1↑2↑3↑4↑5↑6+1	"Published Nov 6, 2026" ; "Team Engineering" ; (intro) "Søren Kierkegaard wrote that life can only be understood backward, but it must be lived forward." ; "Agents capture much of their work in logs or transcripts. Researchers at Letta and UC Berkeley proposed sleep-time compute as a way for models to use idle time between interactions to reorganize their existing context. The same idea can help agents organize their memory by reflecting on logs to fix mistakes or produce new notes and representations. This process is often called dreaming" ; (Memory notes) "By “agent memory,” we mean knowledge that Claude gains from experience […] and that it retains between sessions, saved as context it can read later. It excludes other ways models can learn or manage context, such as changes to their weights […], in-context learning (knowledge gained within a single conversation)" ; "But how can Claude tell what’s salient in the moment, before it knows what’s coming next?" ; (Memory can be a prediction…) "Claude started each question in a fresh session, with one exception: it could retrieve files from a memory directory that persisted from one session to the next."	https://www.anthropic.com/research/components-of-agent-memory
S2↑1↑2↑3	(intro) "The hippocampus plays a central role in this, capturing experiences in the moment so that memories form quickly and are immediately available. But only some experiences are consolidated into long-term knowledge." ; "Much of that consolidation appears to happen when our brains are offline. As we sleep, and during quiet rest, the hippocampus replays our recent experiences. This is thought to gradually distill those experiences and integrate them into the neocortex—the part of our brains that stores most of what we think of as knowledge." ; "Agent memory can use a similar division of labor." ; (Memory notes) "chemical signals released before, during, or after an experience can mark a memory as important. Marked memories tend to be replayed more often than unmarked ones during offline periods. This process, called replay, helps stabilize and integrate them into long-term memory."	https://www.anthropic.com/research/components-of-agent-memory
S3↑1↑2↑3↑4↑5	(Claude has learned to judge salience) "we gave each model a folder of memory files that it could read from and write to while it played" ; (Dreaming can correct memory) "Mt. Moon, a cave system with its entrance on the west side of the map and its exit on the east. Near the entrance is a trap: a one-way ledge that drops the player onto the route they came from" ; "Claude wrote down (incorrectly) that the PokéCenter was on the east side. When it walked out of the PokéCenter, it announced, “Mt Moon is behind me,” and updated its notes: “Mt Moon DONE.”" ; "Claude checked the note against the session transcript, caught its error, and rewrote it. […] With the corrected memory note, Claude went straight back into Mt. Moon and eventually reached Cerulean City." ; "In the run with the uncorrected note, Claude dropped back onto that earlier route three times. After the third failure, Claude seemed to give up, writing, “I’m going to stop rather than keep looping.” Our setup’s automatic reply instructed it to continue. So it did, all the way back to Pewter City, the town before the mountain […] The run ended without Claude ever reaching Cerulean City."	https://www.anthropic.com/research/components-of-agent-memory
S4↑1↑2↑3↑4↑5	(Dreaming can build new memory representations) "a wiki, a collection of pages written and maintained by Claude that gathers what Claude knows about one subject and links it to other notes. These links effectively provide Claude with a map" ; "a corpus of internal documents—about 350,000 files, roughly 420M tokens—from Stainless, a software company we acquired in May. We created an evaluation set of 25 questions about this corpus" ; "Over 128 sessions, we prompted Opus 5 agents to explore the files, write or correct linked wiki pages, merge duplicates, and check claims against the sources."	https://www.anthropic.com/research/components-of-agent-memory
S5↑1↑2↑3↑4↑5↑6+4	"As we scaled dreaming from two sessions to 128, performance on the 25 questions grew from 13% to 48% of graded criteria" ; "In the early rounds of dreaming, Claude’s knowledge is like that of a new employee […] like that of a seasoned employee." ; "When it used both search and the wiki, Claude’s aggregate score rose from 54% to 61% as that wiki grew, compared with 53% when it had access to the search tool alone. Claude’s best score was 48% when it used only the wiki to answer questions. Using search and the wiki also saved about 20,000 tokens per question." ; (figure) "line chart of graded criteria met as wiki generation grows from 2M to 128M dreaming tokens. Wiki plus search tool rises from 54% to 61%, wiki alone rises from 13% to 48%, and the search tool with no wiki is flat at 53%." ; "103 pages and 766 links after 2 sessions and 222 pages and 12,893 links after 128 sessions"	https://www.anthropic.com/research/components-of-agent-memory
S6↑1↑2↑3↑4↑5↑6	"we used the Harvey LAB benchmark along with a collection of documents about a merger. The benchmark gives Claude a set of documents and asks Claude to generate a legal brief based on their content. The benchmark then grades the brief against a list of criteria." ; "One test asked 24 questions about the merger […], more than half of which needed Claude to find information across several documents, and many of which also needed Claude to use judgment or do math. On those items (red and orange lines), Claude scored 87% when it consulted the wiki, compared with 73% when it used search alone." ; "brief criteria often could be retrieved with a text search from a single document, and the wiki showed no clear gain in performance relative to search. On another LAB task, the wiki grew larger than the documents themselves" ; "Like the mapmakers in Borges’s one-paragraph short story “On Exactitude in Science,” who drafted a map so detailed it was as large as the empire it represented"	https://www.anthropic.com/research/components-of-agent-memory
S7↑1↑2↑3↑4↑5↑6+1	(Dreaming can focus memory on a specific goal) "one task in the LAB benchmark simulates a company acquisition and provides a set of contracts. The brief is graded on whether it correctly describes the contract terms." ; "several agents wrote to a shared table of contract terms at the same time" ; "When we graded a brief generated from the wiki, Claude had included only 41 of the 57 facts (72%). […] because agents had overwritten one another’s work." ; "we ran a loop in which Claude reflected on its failures, proposed a change to the instructions, and rebuilt the wiki. This loop allowed Claude to discover how important it was to follow the contract terms, and it rewrote the instructions" ; "Each contract got its own page in the wiki, written by a single agent, which quoted the terms word for word […] 100% of the criteria could be met, up from 72%." ; (caption) "57 of 57 facts survived."	https://www.anthropic.com/research/components-of-agent-memory
S8↑1↑2↑3	(Thinking backward, looking forward) "Dreaming is one way to fix these mistakes (Figure 8), but it requires compute. Directing Claude to continuously check all its memories might not be practical. It may be important instead to focus dreaming on information that’s likely to change or to carefully tune the cadence of dreaming" ; "Claude needs to use compute to maintain memory representations and make sure they don’t accumulate errors or hallucinations." ; (Acknowledgements) "Written by Lance Martin"	https://www.anthropic.com/research/components-of-agent-memory
S9↑1↑2↑3	Claude Platform release notes, re-read on 08/11/2026 at 02:10 UTC, entry of 06/05/2026: "Dreams for Claude Managed Agents are now available as a research preview. A dream reads an existing memory store alongside past session transcripts and produces a reorganized output memory store with duplicates merged, stale entries replaced, and new insights surfaced." ; "Request access to try it."	https://platform.claude.com/docs/en/release-notes/overview
S10↑1↑2↑3↑4	"Dreams" documentation for Claude Managed Agents, read on 08/11/2026 at 02:10 UTC: "Dreams are in research preview." ; "A dream reads an existing memory store alongside past session transcripts, then produces a new, reorganized memory store: duplicates merged, stale or contradicted entries replaced with the latest value, and new insights surfaced." ; "1 to 100 sessions" ; "The input store is never modified, so you can review the output and discard it if you don’t like the result." ; "Request access"	https://platform.claude.com/docs/en/managed-agents/dreams
R1↑	Kira's calculation, S5: 61% − 53% = 8 points of criteria met out of 100 (best wiki + search score versus search alone)	— (in-house calculation)
R2↑	Kira's calculation, S7: 57 − 41 = 16 graded facts missing from the first brief; 41 / 57 = 71.9%, rounded to 72% by the post	— (in-house calculation)