Anthropic: Claude ‘Dreams’ Took Wiki Scores From 13% to 48%

Anthropic says letting Claude “dream”, reviewing past session logs offline to rewrite its own memory, lifted its score on a 25-question test from 13% to 48% of graded criteria as dreaming scaled from two sessions to 128, according to a research post Anthropic published on November 6. By our arithmetic, that is a 3.7x gain.

The feature behind it, called Dreams, is offered in Claude Managed Agents. Anthropic’s documentation says “Dreams are in research preview on select models” and are billed at standard API token rates for the chosen model.

Business Pill · Data Flywheel

An evergreen Business Engineer explainer on data flywheels. It teaches the general idea only and does not discuss Anthropic’s Dreams feature.

What Anthropic Says Dreaming Does

The documentation describes the feature as a way to “Let Claude review past sessions to curate an agent’s memory and surface new insights”. It says that over many sessions a memory store “accumulates duplicates, contradictions, and stale entries”, and that a dream reads the store alongside 1 to 100 past session transcripts and produces a new, reorganized store.

The input store is never modified, the documentation says, so a developer can review the output and discard it. The research post ties the idea to how sleep consolidates human memory and credits researchers at Letta and UC Berkeley with proposing “sleep-time compute”. It says “We’ve found that dreaming has been a useful tool for helping us give Claude effective memory”.

The Numbers in the Research Post

The 13% to 48% result came from a corpus of about 350,000 files, roughly 420M tokens, from Stainless, a software company the post says Anthropic acquired in May. Over 128 sessions, Opus 5 agents built a linked wiki of the files. In the post’s words, “As we scaled dreaming from two sessions to 128, performance on the 25 questions grew from 13% to 48% of graded criteria”.

With both search and the wiki, the post says Claude’s aggregate score rose from 54% to 61% as the wiki grew, against 53% with search alone. On 24 questions about a merger, Claude scored 87% using the wiki against 73% using search alone, on the items that needed information from several documents, judgment or math.

On a contract task from the Harvey LAB benchmark, a first wiki let Claude include 41 of 57 facts (72%) in a legal brief. After Claude rewrote its own wiki-building instructions so each contract got its own page, the post says 100% of the criteria could be met.

Where the Post Says Memory Went Wrong

The post compares memory folders in a Pokémon Red test. Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge; Fable 5 wrote nine files in two folders, plus a single top-level file, and earned four badges, at about 10,000 steps each.

It also lists limits. On one LAB task the wiki grew larger than the documents themselves, and the post says “Claude had made an extremely detailed wiki that effectively abstracted nothing”. On loading notes into every session, it warns “Injected memories use up valuable tokens, and an incorrect note can be carried into every session”.

The Business Lever: Memory That Belongs to the Account

As we read the documentation, a dream’s output is an ordinary memory store kept in the customer’s workspace, built from that customer’s own sessions. On that reading, the longer an agent runs on one platform, the more of its learned rules sit in a store that platform curates, a switching cost that grows with use.

This extends our coverage of Anthropic’s agent stack: Claude Managed Agents added workflows of up to 1,000 agents in October, and Anthropic’s guidance that CLAUDE.md rules are requests while hooks are enforced this week. Our pieces on agents told to evade a monitor and agents that read the grader covered the evaluation side.

The Structural Read

As we read the research post, Anthropic is presenting memory as something Claude curates offline, not only something it writes during a session.

As we read the documentation, the output of a dream is an ordinary memory store kept in the customer’s workspace, built from that customer’s own sessions.

As we read the limits the post lists, more memory is not always better: one wiki grew larger than its source documents, and injected notes cost tokens in every session.

Three Implications

FOR DEVELOPERS The documentation says a dream never modifies its input store, so the output can be reviewed and discarded.

FOR COSTS The documentation says dreams are billed at standard API token rates and that cost scales roughly linearly with the number and length of input sessions.

FOR THE NUMBERS TO WATCH The post reports Anthropic’s own evaluations; the 13% to 48% figure covers 25 questions on one internal corpus.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Four Intelligence Moats.

The essay describes “The Container Moat” as “built by closed data loops inside a customer’s environment”.

As we read it, a memory store that Claude curates from a customer’s own sessions is that kind of loop. The essay does not discuss Anthropic’s Dreams feature.

What Is Not Established

The research post reports Anthropic’s own evaluations, not independent ones, and the 13% to 48% figure covers 25 questions on one company’s internal corpus. The documentation does not list prices beyond standard token rates, and the research post gives no cost per dream. The post does not say when Dreams leaves research preview.

The Bottom Line

Anthropic says dreaming, an offline pass in which Claude rewrites its memory from past session logs, took a wiki-based test from 13% to 48% of graded criteria as it scaled from two to 128 sessions, a 3.7x gain by our arithmetic. The feature is in research preview in Claude Managed Agents, billed at standard token rates.