Claude Can Now "Dream" to Remember: Anthropic's Agent Memory Research Explained
Claude can now "dream" to remember — and Anthropic just showed the numbers
"Dreaming" is offline reflection between sessions: Claude re-reads its session logs while idle, checks its memory notes against the original transcripts, and rewrites them into better-organised knowledge. Anthropic published the research on 6 November 2026, and the headline result is striking: on a 25-question test drawn from about 350,000 internal documents, Claude's score using only its self-built wiki rose from 13 percent to 48 percent as dreaming scaled from 2 sessions to 128, and using the wiki together with search it rose from 54 percent to 61 percent.
The problem is simple. An agent that works across many sessions needs to remember what it learned, but it cannot know in advance what will matter later. Humans solve this in sleep: the hippocampus replays recent experiences offline and distils them into long-term knowledge. Anthropic borrows the idea directly.
 |
| Illustration created for INORYVA |
What "dreaming" actually does
Agent memory, as Anthropic defines it, is knowledge Claude gains from experience — interactions with users and tasks it has been assigned — that it retains between sessions, saved as context it can read later. It excludes changes to the model's weights, knowledge gained within a single conversation, and compaction, the summaries Claude writes to replace older turns. Within that definition, memory is a prediction: by writing down a rule, the agent predicts that the pattern will be relevant again. The post shows how that skill grows across models. On a task from the Continual Learning Bench, which measures an agent's ability to use memory for long-horizon work, Claude Sonnet 4.6 got as far as generating theories about why an answer was wrong, scattering untested guesses across files. Claude Opus 5 was the first to close the loop: most of its entries included a rule, how it knew the rule was true, answers to earlier questions, and clear warnings about traps it had found.
Steering matters, but not in the way you might expect. When Anthropic told Claude to record only failures, both Opus 5 and Opus 4.7 lost much of the performance boost that memory gave them — for stronger models, freedom to judge salience beat fixed rules set in advance. Steering helped older models most: prompting Sonnet 4.6 to read its memory files before answering, and then to test its theories explicitly, nearly doubled its benchmark score. Sonnet 5 does this on its own, reading its memory before answering 9 out of 10 times, against less than half of the time for Sonnet 4.6.
The wiki experiment: 13 percent to 48 percent
The centrepiece evaluation uses a corpus of internal documents from Stainless, a software company Anthropic acquired in May: about 350,000 files, roughly 420 million tokens. Anthropic built a 25-question evaluation set about this corpus, gave Claude a search tool, and instructed it to build a wiki — a collection of linked pages written and maintained by Claude that gathers what it knows about one subject — using dreaming. Over 128 sessions, Opus 5 agents explored the files, wrote or corrected linked wiki pages, merged duplicates, and checked claims against sources, each session building on the last through a shared log.
| Setup | Score on 25 questions | What it shows |
|---|---|---|
| Wiki alone, after 2 dreaming sessions | 13 percent | A new wiki knows almost nothing |
| Wiki alone, after 128 dreaming sessions | 48 percent | Dreaming compounds: the wiki grows richer |
| Search alone | 53 percent | Search is the stronger single tool |
| Wiki plus search, early wiki | 54 percent | Wiki adds little at first |
| Wiki plus search, grown wiki | 61 percent | The combination beats either alone |
Two details matter. First, search alone at 53 percent beat the wiki alone at 48 percent even after 128 sessions — the wiki is not a replacement for search. Second, using search and the wiki together saved about 20,000 tokens per question, because the wiki acted as a map that let Claude find the right memories faster. Anthropic also tested wikis on legal work using the Harvey LAB benchmark and merger documents. On 24 questions about the merger, Claude scored 87 percent consulting the wiki against 73 percent with search alone. On a contract-terms brief, the wiki first captured only 41 of 57 facts (72 percent), After a loop in which Claude reflected on its failures and rewrote the instructions — giving each contract its own page, quoting terms word for word — the brief met 100 percent of the criteria.
Pokémon Red: 4 badges against 1, with far fewer notes
The most vivid test is Pokémon Red, which Anthropic uses to measure Claude's judgement about what to save. Each model played with a folder of memory files it could read and write. About 10,000 steps in, Claude Sonnet 3.7 had written 413 memory files across 225 folders — a separate file for each battle, item, and character — and earned one badge. Claude Fable 5 wrote nine files in two folders, plus a single top-level file called global.md, and earned four badges. Sonnet 3.7's problem was that the game state changes constantly, so its detailed record kept going out of date; around step 10,000 it spent 17 steps creating six files, deleting nine others, and taking no actions in the game at all. Fable 5 stored only what the game did not already show on screen: strategy, plans, and tested rules,
 |
| Illustration created for INORYVA |
Dreaming can correct memory, not just organise it
A note that is wrong from the start, or that drifts as the environment changes, can be read and trusted by Claude in a later session. Dreaming gives Claude a way to test its notes against the original transcripts. Anthropic shows this with a trap in Mt. Moon, a cave in Pokémon Red: Claude wrongly wrote down that the PokéCenter was on the east side of the mountain, walked off a one-way ledge, and concluded it had overshot west. In the run with the uncorrected note, Claude dropped back onto the earlier route three times, gave up, and walked all the way back to Pewter City without ever reaching Cerulean City. When dreaming was run on the memory folder containing the wrong location, Claude checked the note against the session transcript, caught its error, and rewrote it — and with the corrected note, it went straight back into Mt. Moon and eventually reached Cerulean City. The transcript gave Claude a ground truth to validate its memory against.
The limits Anthropic admits
The post is unusually honest about what does not work yet. Dreaming requires compute, and directing Claude to check all its memories continuously may not be practical — Anthropic suggests focusing dreaming on information likely to change, and tuning how often it runs based on how often memories are accessed. Excess information can overwhelm future sessions: a wiki can grow larger than the documents themselves, and Anthropic compares that to the mapmakers in Borges's story "On Exactitude in Science", who drafted a map as large as the empire it represented. Promoting memories into a new session is risky too: injected memories use up valuable tokens, an incorrect note can be carried into every session, and loading a note does not guarantee Claude acts on it — Anthropic's own documentation warns that longer instruction files reduce adherence. Claude Duet, for example, benefits from memories organised around the person it is helping. For the wider research picture, see our State of AI Report 2026 explainer and our Claude AI drug discovery Anthropic story, the other Anthropic research post in this cluster.
 |
| Illustration created for INORYVA |
Key Takeaways
- Anthropic's "dreaming" is offline reflection between sessions: Claude re-reads logs, checks notes against transcripts, and builds wikis — published as research on 6 November 2026, not a product launch.
- On 25 questions over ~350,000 Stainless files, the wiki alone rose from 13 percent to 48 percent across 128 dreaming sessions; wiki plus search rose from 54 percent to 61 percent, saving about 20,000 tokens per question.
- In Pokémon Red, Fable 5 earned 4 badges with 9 memory files, while Sonnet 3.7 earned 1 badge with 413 files — less, better-chosen memory beat exhaustive memory.
- Dreaming corrected a wrong PokéCenter location by checking the note against the transcript; without the fix, Claude looped three times and never reached Cerulean City.
- The limits are real: dreaming costs compute, wikis can grow larger than their sources, and a wrong injected note can be carried into every session.
Conclusion
Anthropic's dreaming research does not announce a feature anyone can use today, but it puts numbers on a design choice every agent builder faces. Memory is not a filing cabinet to fill, it is a prediction about what the future will need, and the evidence here says that a small, well-chosen, regularly corrected memory beats a large indiscriminate one. For now, the practical takeaway is the one the Pokémon test makes plain: what an agent chooses not to remember matters as much as what it keeps.
Comments
Post a Comment