Anthropic's Dreaming Fixed Mt. Moon and Grew a Stainless Wiki From 13% to 48%. It Is a Research Post, Not a Product.
Anthropic published a research post on November 6 on the components of agent memory, including dreaming, offline sleep-time compute that reorganizes session logs into notes, wikis, and injected context between sessions. The demos: Claude Fable 5 kept 9 memory files and won 4 Pokemon badges in about 10,000 steps where Sonnet 3.7 kept 413 files and won 1; a wiki over Stainless's 350,000-file corpus rose from 13% to 48% of graded criteria across 128 sessions, with search plus wiki at 61% against search alone at 53%; a legal contract wiki went from 41 to 57 of 57 facts after an instruction rewrite. Every number is Anthropic's, one wiki grew larger than its source documents, and nothing here is a Claude product, SKU, or price. We read the post.
By T.W. Ghost
Anthropic published a research post today, November 6, 2026, called "The components of agent memory." The idea it names is dreaming: offline compute that runs between sessions and reorganizes what an agent logged into notes, wikis, and a small block of context injected at the start of the next run. It descends from the sleep-time compute line of work, and the post walks through three demonstrations. None of them is a product. There is no Claude memory SKU, no pricing, and no announcement that Claude does this in production. Read it as design notes for anyone building an agent harness, with the labels attached.
The demos are vivid, which is why they will travel without the labels. In Pokemon, Claude Fable 5 played about 10,000 steps keeping nine memory files in two folders plus a global note and won four badges; Claude Sonnet 3.7 kept 413 files in 225 folders and won one. At Mt. Moon, dreaming corrected a wrong note about where the Pokemon Center was; the uncorrected version looped and never reached Cerulean. On Stainless's internal corpus, about 350,000 files and 420 million tokens, a wiki built over 128 sessions rose from 13% to 48% of graded criteria. And on Harvey's legal-contract benchmark, a contract wiki went from 41 of 57 facts to 57 of 57 after the instructions were rewritten to one agent per contract, quoting terms. Every one of those is Anthropic's evaluation, graded by Anthropic. We read the post. Here is what held up.
The Three Demos, by Anthropic's Count
| Demo | Result (vendor) |
|---|---|
| Pokemon, Claude Fable 5 | About 10,000 steps; 9 files, 2 folders, plus a global note; 4 badges |
| Pokemon, Claude Sonnet 3.7 | About 10,000 steps; 413 files, 225 folders; 1 badge |
| Mt. Moon | Dreaming corrected a wrong Pokemon Center location; the uncorrected note looped and never reached Cerulean |
| Stainless corpus wiki | About 350,000 files, about 420M tokens; 13% to 48% of graded criteria over 128 sessions |
| Search plus wiki | 61% aggregate, up from 54% as the wiki grew; search alone 53%; wiki alone peaked at 48%; about 20,000 tokens saved per question |
| Harvey LAB contract wiki | 41 of 57 facts to 57 of 57 after rewriting instructions to one agent per contract, quoting terms |
| The Borges warning | One contract wiki grew larger than the source documents |
| Salience rules | "Record only failures" reduced the memory benefit for Opus 5 and Opus 4.7 on CL-Bench |
| Product status | Research; no GA, no SKU, no price |
Memory hygiene beat memory volume, in Anthropic's own run. Nine files and four badges against 413 files and one badge is the cleanest line in the post, and it is a comparison between two Claude generations on one game, scored by the lab that makes both.
The wiki is a map, and the map lost to map plus territory. Wiki alone reached 48%. Search alone was 53%. Search with the wiki reached 61%. The organized memory helped most as an index into the raw documents, not as a replacement for them.
Borges is on the page. Anthropic reports that one legal wiki grew larger than the contracts it summarized. A memory system that abstracts nothing is a second copy of the data with a new failure mode.
Rules that help weak models can hurt strong ones. Telling the agent to record only failures cut the memory gain for Opus 5 and 4.7. The harness setting is not model-independent.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Research post published Nov 6, 2026 | Verified (Anthropic research) |
| Dreaming as offline memory organization between sessions | Vendor framing |
| Pokemon file counts and badges; Mt. Moon correction | Vendor |
| Stainless 13% to 48%; search plus wiki 61% vs 53%; about 20,000 tokens saved | Vendor |
| Harvey LAB 41 to 57 of 57 | Vendor |
| One wiki grew larger than its sources; salience rule hurt stronger models | Vendor |
| A Claude product, SKU, or price for dreaming | None (verified absence) |
What To Do
Use it as a design document if you build agents: the split between notes, wikis, and a small injected header is the reusable part, and so is the warning to keep the top layer small and the wiki pages quote-faithful.
Do not ship a "Claude dreams" claim or promise a customer any of these numbers; they are one lab's internal evaluations of its own models, and the product that would make them yours does not exist.
The frontier is elsewhere: the always-on agents that are products, Grok Bot, Dots, and DeepSeek's open prototype, all face the memory problem this post describes, and none of them has published a result like it. Our model comparison has the current standings on the models underneath.
Sources
*Not sure how much memory your agent should keep between sessions? Take the free 2-minute quiz and get matched. Then read about the API feature that keeps a conversation server-side for five minutes: Claude message threads.*