Image: Anthropic — Claude logo (official brand mark). Source: Wikimedia Commons, public domain. Dreaming is a research feature with no product screenshot, so the logo stands in for the official visual.
📌 Key Highlights
- Anthropic published "The components of agent memory" on 6 November 2026 — a design write-up with test results, not a new model launch.
- Agent memory means experience Claude keeps between sessions as readable context. It does not change the model's weights.
- Four layers: transcripts, memory notes, representations such as wikis, and a small layer of injected context.
- "Dreaming" is an offline pass between sessions: it corrects wrong notes, merges duplicates and builds linked wikis.
- Results: with memory, Claude Opus 5.5's score on a database task rose from 0.655 to 0.838; search plus a grown wiki reached 61%, against 53% for search alone.
- Caveat: every figure is Anthropic's own evaluation on small test sets — no independent replication has been published yet.
What happened?
On 6 November 2026, Anthropic's engineering team published a research post, "The components of agent memory", explaining how it designs memory for Claude agents that work across sessions. Written by Lance Martin, it is the first detailed public account of the system behind a feature Anthropic calls "dreaming".
Two things need saying up front, because the name invites misunderstanding. First, dreaming is not new: Anthropic launched it on 6 May 2026, at its Code with Claude developer conference, as a research preview for Claude Managed Agents, its platform for long-running agent work. The 6 November post explains how it works and shares the test results. Second, dreaming is not a new Claude model. It is a scheduled background process that runs between sessions, reviews what an agent did, and rewrites the memory it carries forward. Nothing about the model itself changes.
The idea borrows from neuroscience, and from a research proposal by Letta and UC Berkeley called "sleep-time compute": models can use idle time between interactions to reorganise what they already know, much as the human brain replays the day's experiences during sleep. Anthropic's version is deliberately prosaic — a job that reads files, applies judgement about which facts supersede which, and writes files.
The four layers of Claude's agent memory
Anthropic defines agent memory narrowly: knowledge Claude gains from experience — interactions with users and tasks it is assigned — and retains between sessions as context it can read later. It excludes three things people often mean by "memory": changes to the model's weights during training, learning inside a single conversation, and compaction, the summaries Claude writes to replace older turns.
| Layer | What it stores | How Claude uses it |
|---|---|---|
| Transcripts | The raw record of conversations and activity | A searchable event log — the ground truth dreaming checks notes against |
| Memory notes | Short lessons, rules, goals and plans Claude writes for itself | Read at the start of later sessions to avoid repeating mistakes |
| Representations | Organised structures, such as linked wiki pages | A map Claude can navigate to find the right memory quickly |
| Injected context | A small set of facts preloaded into every new session | Available without searching — kept deliberately small, because a wrong note here rides into every session |
The top layer is small on purpose. Injected notes cost tokens in every session, and loading a note does not guarantee Claude acts on it — Anthropic's Claude Code documentation warns that longer instruction files reduce adherence. Claude Code's auto memory loads only the first 200 lines or 25 KB of a MEMORY.md file as an index of what exists and where to find it. Much of dreaming's work is keeping that top layer small: moving details down the stack and leaving links behind.
Illustrative image: the hippocampus, the brain region Anthropic uses as the analogy for Claude's dreaming — replaying experiences during sleep to decide what to keep. Image: explorebiology, Wikimedia Commons, CC BY 4.0.
What "dreaming" actually does
Dreaming runs between sessions, when the agent is idle. It takes the existing memory store and a batch of session transcripts — Anthropic describes a Dreams API that runs it as an asynchronous job over between one and 100 transcripts — and produces a reorganised memory store. The original transcripts stay untouched as the record it checks against. Three jobs happen in a dream: correct mistakes, build representations, and promote the facts worth loading into every future session.
The correction job is best shown by a test Anthropic ran in Pokémon Red, which it uses as a long-horizon task. Claude had wrongly noted that the PokéCenter, its landmark, was on the east side of Mt. Moon. It walked out, announced "Mt Moon is behind me", and updated its notes to "Mt Moon DONE". It was not behind it. Claude then dropped over a one-way ledge onto the route it came from, three times, and eventually wrote "I'm going to stop rather than keep looping" and walked back to the previous town — the run ended without reaching Cerulean City. When dreaming ran on the same memory folder, it checked the note against the transcript, caught the error and rewrote it. With the corrected note, Claude went back into the cave and reached Cerulean City at action 1,935.
The representation job is the wiki: over many sessions, Claude writes and maintains linked pages about what it knows, merging duplicates and checking claims against sources. The promotion job floats the most useful notes to the top of the stack, so the agent harness can preload them into every new session. Dreaming can also be tuned to a goal: Anthropic gives the example of Claude Duet, an assistant that benefits from memories organised around the person it helps.
Anthropic Claude Dreaming Agent Memory: What the Tests Actually Showed
Anthropic ran three kinds of evaluation. The headline numbers all come from its own post, on small test sets, and none has been independently replicated — read them as evidence the design works in Anthropic's hands, not as a benchmark ranking.
| Test | What changed | Result (Anthropic's own figures) |
|---|---|---|
| CL-Bench 1.0 database task | Fresh sessions, carrying only a memory folder forward | Mean reward rose with memory for every model: Opus 5.5 0.655 → 0.838; Sonnet 5.5 0.644 → 0.833; Opus 5 0.564 → 0.828 |
| Pokémon Red — what to save | Judgement about which memories are worth keeping | Sonnet 3.7 wrote 413 files across 225 folders, 1 badge; Fable 5 wrote 9 files plus a global.md, 4 badges; Opus 5.5 kept a small notes folder and beat the game |
| Wiki test — 25 questions, ~350,000 files | A wiki built by dreaming over 2 to 128 sessions | Wiki-only 13% → 48% of graded criteria; search alone 53%; search plus wiki 54% → 61%, saving about 20,000 tokens per question |
| Merger questions — 24 questions | Mostly needed information from several documents | 87% with the wiki, against 73% with search alone |
Two findings cut against the hype. First, the wiki did not always help: on a legal-brief task whose answers could mostly be retrieved from a single document, it showed no clear gain over search. Second, a wiki can outgrow its own documents — on one task, 10 documents and 115 KB of text became 68 pages, 255 links and 203 KB.
The most striking result came from instructions, not scale. On a contract-terms task, several agents writing into one shared table overwrote each other, and a brief built from that wiki contained only 41 of 57 facts (72%). Anthropic then had Claude reflect on the failures and rewrite the instructions: each contract got its own page, quoting the terms word for word instead of summarising them. The rebuilt wiki met 100% of the criteria. The lesson Anthropic draws is to tell Claude what memory is for, not how to use it — a rule to "record only failures" actually cut the memory gain for the stronger models.
Why it matters for readers in India
Most readers will never touch Claude Managed Agents — dreaming is a research preview for that platform, not a feature in the consumer Claude app. The value of this post is the design, and it is immediately usable by anyone building agents, including the growing number of Indian developers and students working with Claude Code and other agent frameworks.
The practical rules are simple, and Anthropic's tests support each one. Store rules and verified facts, not state you can re-read from the screen — in the Pokémon test, the model that recorded everything fell behind the one that recorded only what the game did not already tell it. Keep the memory your agent loads every session short and verified: Claude Code loads only the first 200 lines or 25 KB of MEMORY.md, so treat that file as an index, not an archive.
There is a wider context, too. Anthropic's models are the subject of METR's finding that AI is now accelerating AI research itself, and an Anthropic model recently proved the Aaronson-Ambainis conjecture with a Lean-checked proof. Long-running agents are becoming the way serious work gets done — and an agent that cannot remember reliably will repeat its mistakes at machine speed. The stakes rise as agents are trusted with consequential work: OpenAI's GPT-6 Cyber, for example, is a model built to find software flaws. Anthropic is candid about what it does not know yet: when Claude should create memory, how to keep it reliably up to date, and how often dreaming should run, since checking every memory continuously would cost too much compute.
Should you care?
FAQs
What is Claude dreaming?
Dreaming is a scheduled, between-sessions process for Claude Managed Agents that reviews past session transcripts and the agent's memory store, then produces a reorganised memory — merging duplicates, correcting errors, and building linked wikis. Anthropic introduced it on 6 May 2026 as a research preview and explained it in detail in its 6 November 2026 engineering post.
Is Claude dreaming available to everyone?
No. It is scoped to Claude Managed Agents as a research preview, so consumer Claude users do not get it directly. Anthropic has published cookbooks for developers who want to try dreaming and build knowledge wikis with it.
Does dreaming change Claude's model or retrain it?
No. Dreaming rewrites only the curated memory layer — the notes and wikis Claude reads at the start of future sessions. It does not change the model's weights, and the original session transcripts remain untouched as the record dreaming checks its notes against.
How does Claude Code's memory work?
Claude Code keeps auto memory in a MEMORY.md file that Claude maintains during the session. It loads only the first 200 lines or 25 KB of that file automatically into a new session, as an index of what it knows and where to find it; deeper material is retrieved only when needed.
Takeaway: Claude's memory is not a bigger context window — it is a filing system with a night shift. Dreaming checks the files, fixes the wrong ones and builds a map, and Anthropic's tests say that is worth more than storing everything. The open questions are cost and cadence, not whether memory matters.
If you build agents, the useful next step is the practical one in this post: keep the memory your agent loads into every session short, verified and organised around a goal — then read Anthropic's full write-up for the diagrams behind the numbers.
Sources
- Anthropic — "The components of agent memory" (6 November 2026)
- Subagentic — "Anthropic maps how Claude keeps agent memory"
- AccessAllGPT — "Anthropic Details a Four-Layer Memory Design for Long-Running Claude Agents"
- ExplainX — "Agent Memory: Anthropic's Dreaming Tests, 13% to 48%"
- PakTechStan — "Anthropic Says Claude 'Dreams' Between Sessions to Fix Its Own Memory"
- FourWeekMBA — "Anthropic: Claude 'Dreaming' Took Wiki Score From 13% to 48%"
Comments
Post a Comment