Takeaway: On November 6, 2026, Anthropic published The components of agent memory, an engineering research post by Lance Martin that explains how Claude keeps knowledge between sessions — and how an offline process Anthropic calls dreaming reorganizes that memory. The post is not a claim that Claude updates its model weights while it “sleeps.” Dreaming reads session transcripts plus an existing memory store, then writes a new, cleaned-up store: duplicates merged, stale or contradicted notes replaced, and new representations (including linked wikis) added. In Anthropic’s headline evaluation, Claude built a wiki over about 350,000 internal files (~420 million tokens) from Stainless, a software company Anthropic says it acquired in May. After 128 dreaming sessions, scores on 25 questions rose from 13% to 48% when Claude used only the wiki; combining the wiki with search reached 61%, versus 53% for search alone, and saved about 20,000 tokens per question.
This is a research explainer about memory design for long-running agents. It is distinct from this morning’s METR assessment of Anthropic’s R&D pace, from Claude API message threads, from Claude Personal Finance, from FlashProve, and from this week’s security write-up of Claude Code settings files leaking secrets on GitHub. Dreaming itself was already documented earlier in 2026 as a research-preview feature for Claude Managed Agents; the November 6 post is the lab’s public evaluation of how memory and dreaming actually behave.
What Anthropic means by memory — and what it does not
Anthropic defines agent memory as knowledge Claude gains from experience — user interactions and assigned tasks — and retains between sessions as context it can read later. The post explicitly excludes three other mechanisms: changes to model weights, in-context learning inside a single conversation, and compaction (summaries that replace older turns so a chat can continue in a new context window).
The analogy the authors use is human memory consolidation. People encode experience quickly, then replay and distill some of it during sleep or quiet rest. Researchers at Letta and UC Berkeley previously described “sleep-time compute” as a way for models to reorganize context between interactions. Anthropic applies a similar split: write notes while work is happening, then dream offline to clean, correct, and restructure those notes before the next session.
A memory note is also a prediction. When Claude writes down a rule or a verified fact, it is betting that a future session will need that information without repeating the original work. A wrong prediction is dangerous precisely because later sessions may trust the note.
Stronger models got better at judging what to save
Anthropic reports several evaluations rather than a single leaderboard. On Continual Learning Bench (CL-Bench) 1.0, Claude answered a series of database questions in fresh sessions, with only a persistent memory directory carrying over. After each answer it learned the correct solution and could write a note for later questions.
Different model generations used that loop differently:
- Claude Sonnet 4.6 generated theories about mistakes but left a scattering of files full of untested guesses.
- Claude Opus 4.7 consolidated notes into a single reference and started marking which facts it had checked.
- Claude Sonnet 5 marked answers as confirmed after the grader checked them and wrote corrections when it had been wrong.
- Claude Opus 5 was the first of the set to close the loop: most entries included a rule, how Claude knew the rule was true, answers to earlier questions, and warnings about traps.
- Claude Opus 5.5 and Claude Sonnet 5.5 also closed the loop. With memory, both scored about the same as Opus 5, though their no-memory baselines were slightly higher.
The lab also found that telling stronger models exactly what to write can backfire. A rule to “record only failures” removed much of the performance boost memory had given Opus 5 and Opus 4.7. Anthropic’s usual guidance for frontier Claude models is simpler: create notes when a person asks or when something is clearly worth remembering, and leave the salience judgment to Claude. Older models needed more explicit prompts — for example, telling Sonnet 4.6 to read its memory files and test its theories nearly doubled its benchmark score.
Pokémon as a test of salience, not just storage
To test what Claude thinks is worth remembering, Anthropic compared memory folders from Claude Sonnet 3.7 and Claude Fable 5 about 10,000 steps into Pokémon Red:
| Model | Memory folder around step 10,000 | Progress reported |
|---|---|---|
| Claude Sonnet 3.7 | 413 memory files across 225 folders | One badge |
| Claude Fable 5 | Nine files in two folders, plus a top-level global.md | Four badges |
| Claude Opus 5.5 | A small folder tracking the current goal and next steps | Beat the game, per Anthropic |
Sonnet 3.7 treated almost everything as memorable — separate files for battles, items, and characters — and then spent turns filing memories instead of playing. Around step 10,000 it started a fourth folder of “master” files and, over the next 17 steps, created six files and deleted nine others without taking an in-game action. Fable 5 used memory for things the game did not already display on screen: strategy, plans, and tested rules. Its global.md held the current goal, a gym plan, and the model’s own rules for using memory.
The contrast is the point Anthropic wants readers to take: more memory is not better memory. Notes that duplicate state the agent can reread every turn go stale immediately. Notes that capture a hard-won rule disappear unless someone writes them down.
Dreaming can correct a note that would otherwise be trusted
Because Claude writes memory in the moment, a bad note can be reused confidently later. Dreaming gives Claude a second look against the original transcript. Anthropic’s example is a Pokémon navigation error in Mt. Moon. Claude used a PokéCenter as a landmark and wrote down the wrong side of the mountain for that building. It then concluded “Mt Moon DONE,” walked off a one-way ledge, and never reached Cerulean City in that run — dropping back onto the earlier route three times before appearing to give up.
When Anthropic ran dreaming on the folder that contained the wrong location, Claude checked the note against the transcript, rewrote it, and — with the corrected memory — went back into the cave and eventually reached Cerulean City. The uncorrected-memory run and a no-memory run both failed to get past the mountain, according to the post.
Wikis: 13% to 48% on a 350,000-file corpus
Dreaming can also build new representations. One of those is a wiki: pages Claude writes and maintains about a subject, with links that act as a map through a growing knowledge base. Anthropic tested this on a corpus of about 350,000 internal Stainless files (roughly 420 million tokens) and a 25-question evaluation set.
Opus 5 agents spent 128 sessions exploring the files, writing or correcting linked pages, merging duplicates, and checking claims against sources. They used a shared log to coordinate:
| Setup | Result Anthropic reports | What it does not prove |
|---|---|---|
| Wiki only, after 2 dreaming sessions | 13% of graded criteria | The wiki is not yet a substitute for search |
| Wiki only, after 128 sessions | 48% | Still below search; the wiki is incomplete |
| Search tool alone | 53% | No structured memory map |
| Search plus the grown wiki | 61%; about 20,000 fewer tokens per question | Gains are task-dependent, not universal |
Wikis helped most when questions required combining information across documents. On 24 questions about a merger, more than half of which needed several documents — and many of which also needed judgment or math — Claude scored 87% with the wiki versus 73% with search alone. On some Harvey LAB legal-brief tasks, criteria could be retrieved from a single document with text search, and the wiki showed no clear gain. In another case the wiki grew larger than the source documents themselves, abstracting almost nothing — Anthropic compares that failure mode to Borges’s map the size of the empire.
Instructions change what dreaming keeps
Dreaming can be pointed at a goal. On a LAB acquisition task graded against contract terms, several agents wrote to one shared table at the same time. Only 41 of 57 graded facts (72%) survived into the wiki. After a loop in which Claude reflected on failures and rewrote the wiki-building instructions, each contract got its own page, written by a single agent that quoted terms instead of summarizing them into a contested table. The brief generated from that wiki then met 57 of 57 criteria.
A second dreaming pass also closed gaps a live session left behind. On one task, dreaming found nine links to a missing page and repointed them to an existing page for the same document, then added three new pages including an index.
Anthropic notes a related constraint from Claude Code’s auto memory: Claude does not load an entire memory file into every session. It loads only the first 200 lines or 25 KB of MEMORY.md. That header is an index. Dreaming’s job at the top of the hierarchy is to keep that injected layer small, because a wrong note loaded into every session wastes tokens and can steer later work off course. Longer instruction files, Anthropic’s own Claude Code documentation already warns, also reduce adherence.
Open questions the post leaves on the table
The authors flag design problems that are not solved:
- When should Claude write memory during a task (immediate, but distracting) versus push that work into a later dream (less disruptive, but not instantly available)?
- How often should dreaming run, and which memories are worth the compute to re-check, given that notes can be wrong from the start or drift as the environment changes?
- When is a searchable transcript enough, and when does a maintained representation such as a wiki pay for itself?
- How do you stop representations from accumulating errors — or growing until they are no smaller than the raw files?
Anthropic’s stance is pragmatic rather than triumphant: frontier models have gotten better at judging salience, but no model can know what a future session will need. Dreaming is a way to spend compute looking backward so later sessions can look forward with a cleaner map.
What this is — and what it is not
| Claim | Status | Notes |
|---|---|---|
| Anthropic published “The components of agent memory” on November 6, 2026 | Confirmed | Engineering post by Lance Martin |
| Dreaming updates Claude’s model weights overnight | Not what Anthropic claims | Memory is stored context, not training |
| Dreaming reorganizes an existing memory store using past session transcripts | Confirmed in the post and docs | Input store is left intact; output is a new store |
| 350,000-file Stainless corpus; wiki-only scores 13% → 48% over 128 sessions | Reported by Anthropic | Internal documents; we did not re-run the eval |
| Search plus wiki reached 61% vs. 53% for search alone | Reported by Anthropic | About 20,000 tokens saved per question |
| Wikis always beat search | Not claimed | Some single-document tasks showed no gain; one wiki outgrew its sources |
| This post is the first public mention of dreaming | No | Managed Agents docs already describe dreams as a research preview |
| This is the same story as METR’s November 5 R&D-acceleration report | No | Different question, dataset, and authors |
FAQ
Does dreaming mean Claude trains itself while idle?
No. Anthropic is describing an offline job that reads transcripts and memory files, then writes a reorganized memory store. The post says this is separate from changing model weights.
Is this a new product launch?
Not in the sense of a first announcement. Dreaming is already a research-preview capability for Claude Managed Agents, with API docs and a request-access path. The November 6 post is the research write-up of how memory notes, salience, wikis, and dreaming interact — including new evaluation numbers.
Do I need Managed Agents to benefit from the idea?
The productized dream job is documented for Managed Agents. The broader lesson — write sparse, tested notes during work, then spend idle compute cleaning them against transcripts — also shows up in Claude Code’s auto-memory design, where only a small header of MEMORY.md is injected into each session.
Why did a smaller memory folder beat a huge archive in Pokémon?
Because the game state changes constantly. Sonnet 3.7 stored details the screen already showed, so those notes went stale and then competed for attention. Fable 5 stored plans and rules the game would not remember for it. Anthropic’s argument is that salience beats volume.
Should teams treat the 13% to 48% wiki result as a general benchmark?
No. It is one internal 25-question set over one company’s document corpus, reported by Anthropic. The same post shows tasks where a wiki added little and a case where the wiki became larger than the files it was meant to summarize.
Sources: Anthropic — The components of agent memory (Lance Martin, November 6, 2026); Claude Docs — Dreams; New in Claude Managed Agents: dreaming, outcomes, and multiagent orchestration. Figures and examples above are taken from those pages. We did not independently rerun Anthropic’s evaluations or inspect the Stainless corpus.