What dreaming actually is

Strip the metaphor and dreaming is a batch job. Between sessions, Claude re-reads transcripts of what it did, compares them with the notes it saved, corrects notes that do not match, merges duplicates, and sometimes rebuilds the notes into a linked wiki, a set of pages that point at each other so a later session can navigate them. No weights change. Nothing is learned in the training sense. The model is editing its own notebook, with the original record open beside it.

That distinction matters because the notebook is where agents quietly fail. A note written in one session is read, and trusted, in the next. Anthropic gives a concrete case. Playing Pokémon, Claude saved the location of the PokéCenter at the entrance to Mt. Moon on the wrong side of the mountain. Acting on that note, it walked off a one-way ledge and landed back on the route it had come from, three times. After the third failure, Claude wrote that it would stop rather than keep looping. The test harness instructed it to continue, so it returned to Pewter City to heal and restock. The run eventually ended without reaching Cerulean City. A dreaming pass then checked the note against the session transcript, caught the error and rewrote it. With the corrected note, Claude went back into the cave and reached the city. One case proves a mechanism, not a rate. Anthropic does not say how often dreaming catches errors like this, or how often it introduces new ones.

One Pokémon Run Shows Two Very Different Memory Strategies

The post's most quoted contrast comes from Pokémon. Given a folder to write memory in, Sonnet 3.7 saved almost everything: 413 files across 225 folders by about step 10,000, including a separate file for individual battles and items. Much of that record was game state that changes constantly, so it went stale as fast as it was written, and the model spent growing amounts of time searching its own archive. At one point it spent 17 steps creating six files and deleting nine, and took no actions in the game. It earned one badge. Fable 5 kept nine files in two folders, plus a single top-level file holding its current goal, its plan and its own rules for what to save. It earned four badges. Opus 5.5 kept a similarly small folder and finished the game.

Read that as Anthropic presents it, with the limit stated in its own text: one run per model. A single run cannot separate a memory strategy from luck, route choices or the model itself, and Fable 5 is also simply a newer model than Sonnet 3.7. The narrower observation survives: Sonnet 3.7 repeatedly stored changing game state and spent substantial effort maintaining its archive, while Fable 5 emphasized goals, plans and tested strategies. The comparison illustrates the costs of indiscriminate memory writing. It does not establish how much of the performance difference was caused by memory strategy rather than model capability or variation between runs.

The database task in CL-Bench 1.0 points the same way, with a sharper edge for anyone writing agent instructions. Newer models used memory to build a learning loop: record a failure, test a theory about it, keep the rule that survived. Opus 5 was the first to close that loop fully. But when Anthropic told the models to record only failures, both Opus 5 and Opus 4.7 lost much of the benefit memory had given them. A specific, plausible instruction made strong models worse. Weaker models needed the opposite: telling Sonnet 4.6 to read its memory before answering and to test its theories nearly doubled its score, a habit Sonnet 5 already has, reading its memory first in nine out of ten cases without being told.

The wiki numbers, read carefully

Dreaming's second job is building representations, and the wiki experiment is where the costs show. Anthropic fed Claude about 350,000 internal files from Stainless, a company it acquired in May, roughly 420 million tokens, and had agents build a linked wiki over repeated dreaming sessions, then answer 25 questions from it.

Across 25 evaluation questions, wiki-only performance increased from 13% to 48% of graded criteria as Anthropic expanded the wiki from two to 128 dreaming sessions. With both search and wiki available, performance rose from 54% to 61% over the same expansion. Search alone scored 53%. The final combined system therefore exceeded the search-only baseline by eight percentage points, while the incremental improvement from expanding the combined system's wiki was seven points. It also means that after 128 sessions of offline work, the wiki on its own still missed more than half of what the graders asked for. Search plus wiki also saved approximately 20,000 tokens per evaluation question, according to Anthropic. That is an inference-time saving, not a demonstrated reduction in total cost. Building and maintaining the wiki required up to 128 dreaming sessions, and the post does not provide a dollar-cost comparison or establish a break-even point.

In a separate evaluation involving 24 merger-document questions, Anthropic found a larger benefit on items requiring information from several documents or additional reasoning: 87% of graded criteria with the wiki, compared with 73% using search alone. That is a 14-percentage-point difference in a different task, not a breakdown of the 25-question Stainless evaluation. On single-document questions it added nothing measurable, and in one case the wiki grew larger than the source documents, which is a summary that has stopped summarizing. In another documented repair pass, dreaming corrected nine links to a nonexistent wiki page and added three pages, including an index; Anthropic did not report a general link-repair success rate.

Anthropic is open about the remaining problems. Dreaming costs compute, checking every memory continuously is not practical, and how often to run it is an open tuning question. Anthropic warns that injecting incorrect memories into every new session can repeatedly propagate errors. As one implementation example, Claude Code auto memory loads only the first 200 lines or 25 KB of MEMORY.md, using that header as an index to other stored information. The post does not establish that this limit applies to every Claude agent or memory system. For builders, that is the practical content of the post. Keep the always-loaded layer tiny, let the model judge what is worth saving, verify saved notes against transcripts on a schedule, and expect to pay for the verification. The word dreaming is optional.

Sources

  1. [1] Anthropic, "The components of agent memory," research post, 6 November 2026, read in full this run: the definition of agent memory and its exclusions; dreaming and the Letta / UC Berkeley sleep-time compute credit; the CL-Bench 1.0 results including the "record only failures" reversal and Sonnet 4.6's near-doubling under explicit steering; the Pokémon comparison (one run each, 413 files and 225 folders and one badge for Sonnet 3.7, nine files plus global.md and four badges for Fable 5, Opus 5.5 finishing the game); the Mt. Moon correction case; the Stainless wiki experiment (about 350,000 files, roughly 420M tokens, 25 questions, 13% to 48% wiki-only from two to 128 sessions, 61% search-plus-wiki against 53% search alone, about 20,000 tokens saved per question, 87% versus 73% on multi-document merger questions); the contract-terms rebuild (41 of 57 facts, then 100%); the missing-page repair (nine links repointed, three pages added); the Claude Code auto-memory limit (first 200 lines or 25 KB); and the stated open questions on compute and dreaming cadence.Read source
  2. [2] Anthropic on X (@AnthropicAI), 6 November 2026, as quoted in secondary coverage this run: "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."
  3. [3] Interesting Engineering, "Can Claude dream now? Anthropic explores human-like machine memory," published 7 November 2026: secondary confirmation of the post's date and the Pokémon and dreaming details; its headline framing ("dream now", "eerily close to human thinking") is the circulating claim the Claim Check social copy addresses.Read source