Anthropic published an engineering post on 6 November 2026, “The components of agent memory”, that describes how Claude keeps knowledge between sessions and a process it calls dreaming, in which Claude reviews logs and transcripts to fix mistakes and reorganise its memory notes.
In one of the evaluations it reports, Anthropic says that as it scaled dreaming from two sessions to 128, Claude’s performance on 25 questions about a large document corpus “grew from 13% to 48% of graded criteria”.
Business Pill · AI Agents Explained
A short explainer on AI agents. It teaches the general idea only and does not discuss Anthropic’s memory post.
The key insight: As we read the post, Anthropic’s results put memory design on the harness side of the agent stack: the same Claude models did better or worse depending on how memory was organised, checked and loaded, not only on which model was used.
The post, written by Lance Martin, defines agent memory as “knowledge that Claude gains from experience” that it keeps between sessions as context it can read later. It excludes changes to model weights, in-context learning and compaction. Dreaming, it says, borrows from research by Letta and UC Berkeley on sleep-time compute, in which models use idle time to reorganise context.
Share of graded criteria met on 25 questions about a corpus of about 350,000 internal documents, as stated in Anthropic’s post of 6 November 2026: 13% with a wiki built over two dreaming sessions, 48% after 128 sessions, 53% with search alone and 61% with search plus the grown wiki.
What Was Posted on X
Anthropic found Claude Sonnet 3.7 getting stuck managing its own memory while playing Pokรฉmon.
โOver the next 17 steps, it created six files, deleted nine others, and took no actions in the game itself.โ
In one run, it wrote 413 files and earned one badge. A stronger model kept just nine files and earned four badges in its run.
A post at 23:02 UTC on 6 November by @ChrisGPT quotes Anthropic’s memory post on Claude Sonnet 3.7 playing Pokรฉmon: over 17 steps it “created six files, deleted nine others, and took no actions in the game itself.” Anthropic’s post places that stretch around step 10,000. The 413-file and nine-file runs the post mentions match the figures reported in this piece.
The Wiki Results
Anthropic says it tested wikis, linked pages Claude writes and maintains, on a corpus of about 350,000 files, roughly 420M tokens, from Stainless, a software company it says it acquired in May. Over 128 sessions, Opus 5 agents explored the files, wrote and corrected pages, merged duplicates and checked claims against the sources.
With the wiki alone, Claude’s best score was 48%. With search and the wiki together, Anthropic says the aggregate score rose from 54% to 61% as the wiki grew, compared with 53% with search alone, and that search plus the wiki saved about 20,000 tokens per question.
On 24 questions about a merger, more than half needing information from several documents, Anthropic says Claude scored 87% with the wiki against 73% with search alone. On one legal-brief task from the Harvey LAB benchmark, it says the wiki showed no clear gain, and on another the wiki grew larger than the documents themselves.
What Each Model Saved
In Pokémon Red, Anthropic compared memory folders at about 10,000 steps. It says Claude Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge, while Claude Fable 5 wrote nine files in two folders, plus a single top-level file, and earned four. It adds that Opus 5.5 kept a small folder of notes and beat the game.
On Continual Learning Bench 1.0, Anthropic says Claude Opus 5 was “the first of these models to close the loop”, writing rules with how it knew they were true. It says Sonnet 5 read its memory before answering 9 out of 10 times unprompted, against less than half the time for Sonnet 4.6, and that prompting Sonnet 4.6 to read its memory and test its theories nearly doubled its score.
Rules and Corrections
Anthropic says telling stronger models to “record only failures” cost both Opus 5 and Opus 4.7 much of the gain memory gave them. On one acquisition brief, it says a wiki built by several agents writing to a shared table kept 41 of 57 graded contract facts (72%). After Claude rewrote its own instructions so that one agent quoted each contract word for word, Anthropic says 100% of the criteria could be met.
The post also describes a Pokémon run in which dreaming caught a wrong note about a landmark’s location: with the corrected note, Claude went back into the cave and reached the next city, while the run with the uncorrected note never got past the mountain. It notes that Claude Code’s auto memory loads only the first 200 lines or 25 KB of MEMORY.md into each session.
Our Numbers
By our arithmetic, Fable 5 used 9 files against Sonnet 3.7’s 413, about 46 times fewer, while earning four times as many badges. Scaling dreaming from two to 128 sessions added 35 points on the wiki-only questions. On the same day, Meta said its Muse models used critical context correctly in 90% of tasks, up from 76%, on its own evaluation; the two companies used different tests, so the figures are not comparable.
As we read it, the clearest gain came from combining tools: Anthropic says search plus a grown wiki reached 61% on its 25 corpus questions against 53% for search alone, while saving about 20,000 tokens per question.
As we read the Pokémon comparison, less memory went further: Anthropic says Fable 5 wrote nine files and earned four badges, while Sonnet 3.7 wrote 413 and earned one.
As we read the rules test, Anthropic’s finding cuts against heavy prompting for strong models: it says telling Opus 5 and Opus 4.7 to record only failures cost them much of the gain memory gave them.
Anthropic, The components of agent memory, 6 November 2026
“Dreaming gives Claude an opportunity to look backward and use compute to organize its memory, correct mistakes, and maintain representations that allow future sessions to access memory more readily”
Three Implications
FOR TEAMS BUILDING AGENTS Anthropic says it helps to tell Claude what memory is for, but not how to use it, and that steering helped older models the most.
FOR MEMORY LOADED INTO EVERY SESSION Anthropic says the top layer of the memory hierarchy has to stay relatively small, because injected memories use tokens and an incorrect note can be carried into every session.
FOR THE COST Anthropic says dreaming requires compute and that directing Claude to continuously check all its memories might not be practical.
The analysis says: “The surface, not the model, governs autonomy.”
As we read it, the memory folders, wikis and loading rules Anthropic describes sit in the harness around the model, which sets what an agent can draw on in each session. The analysis does not discuss Anthropic’s memory post.
What Is Not Established
The results come from Anthropic’s own evaluations, several of them on single runs or small question sets, such as 25 corpus questions and 24 merger questions, and the post does not report outside replication. It lists open questions, including when to create memory and how to keep it up to date, and says dreaming requires compute. We did not contact Anthropic. The 46 times and 35-point figures are our arithmetic.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Anthropic says dreaming, in which Claude reviews its own logs to correct and reorganise memory, lifted a wiki-only score from 13% to 48% over 128 sessions, and that stronger models did better when left to judge what to save. The post closes: “we still have a lot to learn about designing Claude’s memory systems”.
A note on sourcing. All results are from Anthropic’s engineering post “The components of agent memory” by Lance Martin, published 6 November 2026, and come from Anthropic’s own evaluations. The 46 times and 35-point figures are our arithmetic. We did not contact Anthropic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
