Anthropic has shown how Claude can learn from experience without being retrained, by letting the assistant review its own notes during quiet moments between sessions. In research published on 6 November, the company describes a process it calls dreaming, in which Claude checks its stored memories against the original session transcripts, corrects mistakes, and builds linked summaries that later sessions can use. The tests show the approach can lift performance sharply on long-running tasks, and that keeping fewer, better-chosen memories can beat keeping everything.
What dreaming actually does
Anthropic defines agent memory as the knowledge Claude gains from experience and keeps between sessions, saved as context it can read later. It is distinct from training, which changes the model’s weights, and from summaries that replace older turns inside a single conversation.
The idea borrows from sleep-time compute, proposed by researchers at Letta and UC Berkeley, in which a model uses idle time between interactions to reorganise what it already knows. Anthropic calls that offline review dreaming, and says it has become a useful tool for giving Claude effective memory.
Less memory, more progress
The clearest test came from Pokémon Red, which Anthropic used to judge what Claude chooses to save. After about 10,000 steps, Sonnet 3.7 had written 413 memory files across 225 folders, nearly 978,000 characters in all, and earned one badge. Fable 5 wrote nine files in two folders, plus a single top-level file, and earned four badges.
The difference was judgment, not effort. Sonnet 3.7 recorded everything, including facts it could read off the screen, and its notes kept going out of date as the game changed. Fable 5 saved only what the game did not show: strategies, plans and tested rules. Opus 5.5 took the same approach, kept a set of notes, and beat the game.
A second test used the Continual Learning Bench, where Claude answers database questions across fresh sessions and may keep a memory folder between them. Sonnet 4.6 got as far as guessing why answers were wrong. Opus 5 was the first to close the loop, recording rules, how it knew they were true, and warnings about traps.
Dreaming fixes mistakes
Dreaming’s sharpest demonstration came when Claude misremembered where a Pokémon Center sat near Mt. Moon. It walked out, declared the mountain done, and dropped off a one-way ledge back onto the route it came from. The run kept looping, and Claude eventually wrote that it was going to stop rather than keep going.
Run through dreaming, Claude checked the note against the session transcript, caught the error, and rewrote it. With the corrected note it went straight back into Mt. Moon and reached Cerulean City. The transcript gave Claude a ground truth to validate its memory against.
From notes to wikis
Dreaming can also build new representations of what Claude knows. In one experiment, Anthropic gave Claude about 350,000 internal documents from Stainless, a software company it acquired in May, and asked it to build a linked wiki. Over 128 dreaming sessions, its score on 25 evaluation questions rose from 13 percent to 48 percent of the graded criteria.
The wiki helped most alongside search: using both, Claude’s aggregate score rose from 54 percent to 61 percent, against 53 percent with search alone, and dreaming saved about 20,000 tokens per question. On a legal task, a wiki built from merger documents answered cross-document questions at 87 percent accuracy, against 73 percent for direct search.
Anthropic also found that telling Claude what memory is for works better than telling it how to use it. A rule to record only failures stripped much of the benefit for the stronger models, while prompting Sonnet 4.6 to read its memory and test its theories nearly doubled its benchmark score.
Dreaming can also promote key notes into the top of the memory hierarchy, so they load into every session automatically. That power needs care: an incorrect note carried into every session costs tokens and can mislead Claude, so the top layer has to stay small.
Our Read
This is Anthropic thinking out loud about the problem every serious agent builder now faces: a model that forgets between sessions cannot do long work, and a model that remembers everything drowns in its own notes. The Pokémon numbers make the point brutally — 413 files and one badge, against nine files and four.
The lesson is not that Claude should remember less in general, but that salience is a skill, and the newer models are visibly better at it. Opus 5.5 and Sonnet 5.5 closed the learning loop that stumped their predecessors, and Sonnet 5 reads its memory unprompted nine times out of ten.
The timing matters too. Memory is where the agent race is now being fought: OpenAI has built persistent memory into ChatGPT, Google is wiring Gemini into its apps, and every coding agent lives or dies by what it carries between sessions. Anthropic’s answer is distinctive — not a bigger context window, but a maintenance routine that curates, corrects and prunes.
None of this is a product yet. Dreaming is a research method and a set of cookbooks, not a feature users can switch on, and Anthropic is candid that it costs compute and that wrong notes can poison later sessions. The honest read is that this is infrastructure thinking, published early, and a signal of where Claude is heading.
Why It Matters
Agents that cannot remember cannot do long work. Every coding assistant, research agent and personal helper that runs across days rather than minutes needs memory it can trust. Today most of them either start cold each session or drag an ever-growing pile of notes behind them.
Anthropic’s experiments show a third path is workable: let the model decide what mattered, then let it check its own homework while idle. If that approach reaches Claude’s products, long-running agents could get steadily more reliable instead of steadily more confused.
Background
Anthropic has been pushing Claude toward longer, more autonomous work all year, from the Opus 5.5 and Sonnet 5.5 launches in September to the Personal Finance beta that asks Claude to reason over two years of transactions. Memory is the missing piece in that push, and this post is the company’s account of how it thinks the piece should work.
Source: Anthropic’s research post and Interesting Engineering’s report.
0 Comments