Anthropic published “The components of agent memory” on November 6, 2026, an engineering post by Lance Martin on how Claude should remember work from one session to the next. The company separates two jobs: writing notes while a task is underway, and an offline pass — often called dreaming — that replays transcripts, corrects mistakes, merges duplicates and builds maps of what the agent already knows. The sharpest public number is simple. On about 350,000 internal files totaling roughly 420 million tokens, wiki-only accuracy on 25 questions rose from 13% to 48% of graded criteria as dreaming scaled from 2 sessions to 128. This is a research result about memory design, not a replacement for the builder tutorials already covered in Developers Share Claude Code Agent Tutorials After Cherny’s Playbook.
Quick Take
- The idea: Anthropic treats agent memory as notes Claude writes from experience and rereads later — not weight updates, not in-context learning, and not compaction summaries.
- Dreaming: Between sessions, Claude can replay logs to fix wrong notes, build wiki-style linked pages, promote a small set of notes into the next session, and fill gaps a single chat left behind.
- The headline eval: On ~350,000 Stainless files (about 420 million tokens), wiki-only scores rose from 13% at 2 dreaming sessions to 48% at 128. Wiki plus search rose from 54% to 61%, versus 53% for search alone, and saved about 20,000 tokens per question.
- The catch: Wrong notes can be trusted in later sessions; some wikis grew larger than the source files; and dreaming costs compute. Anthropic says no model can really know what a future turn will need.
What happened
The post went up on November 6. Anthropic’s official account summarized it the same day: “Memory is what allows Claude to learn from experience,” and dreaming helps Claude “organize what it’s learned, correct its mistakes, and find what it needs more easily the next time.”
Anthropic defines agent memory narrowly. It is knowledge Claude gains from users and assigned tasks and keeps between sessions as context it can read later. The post explicitly excludes changes to model weights, in-context learning inside a single conversation, and compaction (summaries that replace older turns so work can continue in a new context window).
The biological analogy is deliberate. The hippocampus captures experiences quickly; much of the consolidation into longer-term knowledge appears to happen offline, during sleep and quiet rest. Researchers at Letta and UC Berkeley had already proposed “sleep-time compute” as a way for models to reorganize context between interactions. Anthropic applies that idea to memory: reflect on logs, fix mistakes, and produce new notes and representations when the agent is not serving a user.
Background: notes are predictions about the future
A memory note is, Anthropic argues, a prediction that a future session will need the same fact or rule. Writing down a verified answer predicts that the next agent should not have to double-check it. The hard part is judging salience before the agent knows what comes next.
On Continual Learning Bench (CL-Bench) 1.0, Claude answered a series of database questions in fresh sessions, with only a persistent memory directory carried over, and was told the correct answer afterward. Different models used that loop differently. Sonnet 4.6 generated untested theories. Opus 4.7 consolidated notes and marked what it had checked. Sonnet 5 confirmed answers after the grader and wrote corrections. Opus 5 was the first of the compared models to close the loop, storing a rule, how it knew the rule was true, earlier answers and warnings about traps. Opus 5.5 and Sonnet 5.5 also closed that loop.
Steering older models helped, and steering stronger ones could hurt. Telling Opus 5 and Opus 4.7 to “record only failures” removed much of the boost memory had given them. Prompting Sonnet 4.6 to read its memory and test its theories nearly doubled its score; Sonnet 5 already read its memory before answering 9 times out of 10, without that prompt. For frontier Claude models, Anthropic now prefers simple guidance — write a note when asked, or when something is clearly worth remembering — and leaves the judgment to the model.
Pokémon Red made the salience problem visible. At about 10,000 steps, Sonnet 3.7 had written 413 memory files across 225 folders and earned one badge, often storing details the screen already showed. Fable 5 wrote nine files in two folders plus a top-level global.md and earned four badges, keeping strategy, plans and tested rules rather than redundant game-state. Opus 5.5 took a similar sparse approach and beat the game, Anthropic said. The comparison is one run of each model, and later models are also more capable, so the folder sizes are an illustration rather than a controlled ablation.
How dreaming works
Dreaming is the offline counterpart to those in-session notes. Anthropic walks through several jobs it can do.
It can correct memory. In one Pokémon run, Claude wrote down the wrong location for a PokéCenter near Mt. Moon, announced “Mt Moon DONE,” walked off a one-way ledge and never reached Cerulean City — dropping back onto the earlier route three times and later writing, “I’m going to stop rather than keep looping.” When dreaming checked the note against the transcript and rewrote it, Claude went back into the cave and reached the next city. The transcript served as ground truth.
It can build new representations. Anthropic had Opus 5 agents explore the Stainless corpus over 128 sessions, writing or correcting linked wiki pages, merging duplicates and checking claims against sources, using a shared log to coordinate. As dreaming scaled from 2 sessions to 128, performance on 25 questions grew from 13% to 48% of graded criteria when Claude used only the wiki. With both search and the wiki, the aggregate score rose from 54% to 61% as the wiki grew, compared with 53% for search alone. The best wiki-only score was 48%. Using search and the wiki also saved about 20,000 tokens per question.
The wiki helped most when questions required combining several documents or using judgment. On 24 questions about a merger, more than half of which needed information from several documents, Claude scored 87% with the wiki versus 73% with search alone on those harder items. On some Harvey LAB legal-brief tasks, where criteria could be retrieved from a single document, the wiki showed no clear gain. On another LAB task, the wiki grew larger than the documents themselves — Anthropic likens it to the map as large as the empire in Borges’s “On Exactitude in Science.”
Dreaming can also promote a small set of notes to the top of the memory hierarchy so the harness loads them at the start of every session. That injected layer has to stay small: it consumes tokens, a wrong note would repeat in every session, and Claude Code’s own docs warn that longer instruction files reduce adherence. Claude Code’s auto memory already keeps the injected layer small by loading only the first 200 lines or 25 KB of MEMORY.md as an index.
Goal-specific instructions matter. When several agents wrote at once to a shared table of contract terms, only 41 of 57 graded facts (72%) survived. After a loop in which Claude reflected on the failure and rewrote the wiki-building instructions — one agent per contract, quoting terms word for word — 57 of 57 facts survived. A second dreaming pass also closed gaps, for example by repairing nine links to a missing page and adding an index.
| Setup | What Anthropic reported | Limit |
|---|---|---|
| Wiki-only answers, 25 questions | Accuracy rose from 13% to 48% of graded criteria as dreaming scaled from 2 sessions to 128 | Internal corpus from Stainless, a company Anthropic acquired in May; not a public benchmark |
| Wiki plus search vs. search alone | Combined score rose from 54% to 61% as the wiki grew, versus 53% with search alone | The wiki helped most on questions that required combining several documents |
| Corpus size | About 350,000 files, roughly 420 million tokens | Internal documents, not a third-party dataset |
| Token cost | Using search and the wiki saved about 20,000 tokens per question | Dreaming itself consumes extra compute between sessions |
| Pokémon Red memory folders at ~10,000 steps | Sonnet 3.7 wrote 413 files across 225 folders and earned one badge; Fable 5 wrote nine files in two folders plus global.md and earned four badges | One run each; later models also differ in raw capability |
| Merger-contract wiki, Harvey LAB-style task | A shared contract-terms table preserved 41 of 57 graded facts (72%); one page per contract, quoted word for word, preserved 57 of 57 | Anthropic found the better instructions by looping on failures |
Pros and cons
For users
- Pro: Offline cleanup can stop a single wrong note — such as a misplaced landmark — from derailing the next session, and a growing wiki can answer recurring questions with fewer tokens.
- Con: Users still cannot see when a trusted note is stale, and Anthropic warns that injected memories are not guaranteed to be followed.
For developers
- Pro: The post gives a practical split: capture salience in-session, then use idle compute to validate notes against transcripts, build maps and keep the always-loaded layer small. Builders already swapping harness tips after Boris Cherny’s playbook, covered in the earlier Claude Code tutorials story, now have official numbers for the offline half of that design.
- Con: Dreaming is extra compute; parallel agents can overwrite a shared page; and a wiki that merely restates the files adds cost without abstraction.
For companies
- Pro: On a 420-million-token internal corpus, combining a dreamed wiki with search beat search alone and saved about 20,000 tokens per question — a concrete efficiency claim, not just a metaphor.
- Con: The evaluations use Anthropic’s own acquired-company files and internal Pokémon setups. Results may not transfer to messier customer data, and a wrong note loaded into every session becomes an enterprise-wide error.
What X is saying
Anthropic’s own post set the terms. Independent recaps focused on the 13%-to-48% jump and on the risk that memory can poison itself.
Memory is what allows Claude to learn from experience. In a new Engineering Blog post, we share how “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time.
— Anthropic (@AnthropicAI) November 6, 2026 · View on X
Claude’s “dreaming” took a wiki from 13% to 48%. Search alone scored 53%. Anthropic’s engineering post (Nov 6) describes dreaming as an offline pass where Claude replays saved transcripts, fixes wrong notes, builds wiki-style maps and decides which notes load first. It edits memory, not model weights.
— Aimore (@aimorebou) November 7, 2026 · View on X
Anthropic details agent memory in “The Components of Agent Memory”: “dreaming” replays session logs offline to fix mistakes and build wiki-style memory maps. On 350K internal docs, scaling dreaming 2→128 sessions lifted a 25-question eval 13%→48%.
— AI_Pioneer (@Secretof_forest) November 7, 2026 · View on X
Sabriel Atmojo argued that the next bottleneck “isn’t smarter models. It’s whether agents can remember without poisoning themselves,” noting that a bad PokéCenter note sent Claude looping off a ledge until dreaming corrected it.
Critical analysis
Dreaming is not a new model and not extra training. It is housekeeping for files the agent has already written. That is why the Pokémon correction is the clearest anecdote in the post: the transcript was sitting there, and only an offline pass compared the note with what had actually happened.
The 13%-to-48% figure is also easy to overread. It is wiki-only performance on 25 internal questions, and even after 128 sessions it still trailed search (53%) and the combined setup (61%). Anthropic is explicit that the wiki helped most on multi-document questions and did nothing obvious on some single-document legal tasks. The honest reading is that representations plus search beat either one alone — and that unbounded wikis can become another pile of context.
The salience results cut against a popular instinct among agent builders: log everything. Sonnet 3.7’s 413 files were not just clutter; they went stale and stole turns from the game itself. Fable 5’s nine files plus a dashboard-style global.md were more useful because they stored what the screen would not show again. That pairs well with Claude Code’s 200-line / 25 KB cap on injected MEMORY.md: the top of the hierarchy has to stay an index, not an archive.
What’s still open
- When should Claude write memory in-session, where it is immediately available but distracts from the task, versus deferring that work to dreaming?
- How often should dreaming run, and how can it focus on notes that are likely to change rather than rechecking everything?
- Can transcript search replace some learned representations, given that wikis themselves can accumulate errors?
- Will Anthropic ship dreaming as a product feature in Claude Code, Managed Agents or Claude Duet, or keep it as internal research?
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.
This independent, research-based i10x analysis draws on Anthropic’s November 6 engineering post “The components of agent memory,” written by Lance Martin, and on public discussion of that post on X. All benchmark figures are Anthropic’s own measurements on internal evaluations, including documents from Stainless, which Anthropic says it acquired in May. This article covers the research paper; builder tutorials and harness setups are covered separately in Developers Share Claude Code Agent Tutorials After Cherny’s Playbook. More on Anthropic and Claude.
i10x Perspective
The useful shift in Anthropic’s post is not the sleep metaphor. It is the claim that an agent’s notes are hypotheses, and that unused compute should be spent attacking those hypotheses with the original evidence. Teams building long-running coding agents already fight stale CLAUDE.md files and contradictory memories. Dreaming names the missing job: an offline editor that is allowed to delete, merge and demote, not just append. The labs that treat memory hygiene as a first-class system, with caps on what is injected every session, will waste fewer tokens — and fewer runs — on notes that were wrong the day they were written.
