Anthropic's 'dreaming' memory trick lifts Claude agents from 13% to 48%
Anthropic's engineers published research November 6 on how Claude agents remember across sessions — including an offline review pass called "dreaming" that revisits old logs and fixes its own notes. On a roughly 350,000-file corpus, scaling dreaming from 2 to 128 sessions lifted graded performance from 13% to 48%. All figures are Anthropic's own test results, not independently verified.
📊 Scorecard: our Bullish call on $GOOGL is scored from $284.78, the last close before it ran. Its first result comes after the next close.
The News
Anthropic's engineering team published a research post on November 6 dissecting how its Claude agents remember across sessions — the stretches of work an AI assistant does over days and weeks, not just one chat. The system combines stored transcripts, notes the model writes about its own work, linked knowledge structures and a small block of context loaded at the start of each session. The post also describes an offline review pass the team calls "dreaming": the agent revisits old logs while it is not working, and corrects or reorganises its notes.
The headline number comes from a test on a corpus of about 350,000 files. Scaling dreaming from 2 sessions to 128 sessions lifted graded performance from 13% to 48% when the agent answered from the wiki alone — the agent got nearly four times better at the test tasks as it spent more offline time reviewing its own work. When it could also search, the wiki lifted its score from 54% to 61%. Memory also lifted Claude Opus 5.5's mean reward — its average score on the task — from 0.655 to 0.838 on Anthropic's Continual Learning Bench, a test that measures how well an agent keeps learning over time. A mean reward of 0.838 means the agent scored, on average, just over four-fifths of the maximum on the bench's tasks, up from about two-thirds without memory.
Every performance figure here is Anthropic's own test result, not independently verified — the company ran the tests on its own benchmark, and no outside lab has reproduced them yet. Anthropic is private, so there is no stock to buy; Google and Amazon are named only as public read-throughs, the same as in our November 8 Anthropic story — both hold stakes in the company. Google closed Friday at $284.78, up 0.78%, and Amazon at $259.30, down 0.35%.
Dreaming gives Claude a way to test its notes against the original transcripts.
Why It Matters
Memory is the missing piece that keeps AI assistants from being truly useful over time. A chatbot that forgets everything between sessions can answer questions, but it cannot run a project: it cannot remember what it tried, what worked, what a user prefers, or what is still unfinished. Claude already sells to businesses that want an assistant to carry work across weeks; memory is what turns a clever chatbot into a colleague that knows where a project stands. Every assistant has this problem, and the company that solves it first has a real advantage — an agent that remembers is one users come back to.
That is why "dreaming" matters more than the raw percentages. The idea is simple and a little strange: let the agent review its own work while idle, the way people consolidate memories in sleep, and it gets better without any new training data. A jump from 13% to 48% on one test is large, but it is a single test on Anthropic's own benchmark, and the company has every incentive to show its best numbers. Treat the direction as the news and the exact figures as Anthropic's claim, not a settled fact.
Analyst's Take
Our call is Bullish. Agent memory is a genuine bottleneck — every frontier lab is working on it — and Anthropic has published a concrete mechanism with large reported gains rather than a vague promise. The dreaming result, if it holds up outside Anthropic's own tests, points at a way to make agents better without the enormous cost of retraining them.
What to watch: whether other labs reproduce the dreaming gains on their own benchmarks, whether memory features show up in Claude's products for paying users, and whether the November 6 post turns into a product announcement. The risk is the usual one with vendor-reported numbers: impressive on the company's own test, unproven everywhere else. But the idea is cheap to try and the payoff is large, which is why this is worth more than a footnote.
Are you bullish or bearish on this story?
By The Numbers
FAQ
What is an AI agent's memory, in plain terms?
It is how an assistant remembers things between sessions — the stretches of work it does over days and weeks, not just one chat. Anthropic's system combines stored transcripts, notes the model writes about its own work, linked knowledge structures and a small block of context loaded at the start of each session.
What is "dreaming"?
An offline review pass: while the agent is not working, it revisits old logs and corrects or reorganises its notes — the way people consolidate memories in sleep. Anthropic says scaling dreaming from 2 sessions to 128 lifted graded performance on its test from 13% to 48%.
How big was the improvement, exactly?
On a corpus of about 350,000 files, graded performance rose from 13% to 48% as dreaming scaled from 2 to 128 sessions. Memory also lifted Claude Opus 5.5's mean reward — its average score on the task — from 0.655 to 0.838 on Anthropic's Continual Learning Bench, a test of how well an agent keeps learning over time.
Are these numbers independently verified?
No. Every performance figure here is Anthropic's own test result, run on its own benchmark. No outside lab has reproduced them yet, and the company has every incentive to show its best numbers. Treat the direction as the news and the exact figures as Anthropic's claim, not a settled fact.
Anthropic is private — why does this carry tickers?
Because Google and Amazon are named only as public read-throughs, the same as in our November 8 Anthropic story — both hold stakes in the company. Google closed Friday at $284.78, up 0.78%, and Amazon at $259.30, down 0.35%.
👁
Source: Anthropic