Brief IA
›

With Dreaming, Anthropic Aims to Correct Claude's Memories

Models & LLM·Tom Levy·

With Dreaming, Anthropic Aims to Correct Claude's Memories

With Dreaming, Anthropic Aims to Correct Claude's Memories
Key Takeaways
1Dreaming reorganizes existing memory by examining up to 100 previous sessions.
2According to Anthropic, Sonnet 3.7 earned a badge in Pokémon Red, compared to four for Fable 5, in one game each.
3The service requires additional computation and its corrections are not immediate.
Why it matters : Dreaming allows for the verification of notes retained by an agent before they guide its subsequent tasks.

Claude's persistent memory also requires computation and maintenance. Anthropic offers Dreaming to review its agents' notes between sessions, with experimental results that remain limited.

Dreaming reviews up to 100 previous sessions

A computational overhead, deferred corrections, and regular maintenance: agents' memory retains several constraints, acknowledges Anthropic. The company warns that neglected notes or wikis accumulate errors and hallucinations. It also cautions that automatically injecting too many memories consumes available context while passing incorrect notes to subsequent exchanges.

Introduced in May as a research preview within Claude Managed Agents, its cloud platform for AI agents, Dreaming reviews existing memory as well as up to 100 previous sessions. It removes duplicates and outdated items, then generates a reorganized version without altering the original. At launch, only Claude Opus 4.7 and Sonnet 4.6 were compatible.

Anthropic recognizes that several questions remain unanswered. The note-taking process itself diverts the agent from its activity, while subsequent reviews require additional resources and do not yield immediate results.

Pokémon distinguishes two ways to memorize

Anthropic's experiments rely on a mechanism that does not modify the model: Claude writes in a folder, then consults these notes during later sessions. According to the company, the challenge lies in selecting useful information without knowing the upcoming tasks. Its trials favor the selection of rules and strategies rather than an exhaustive history, which quickly becomes cluttered with outdated elements.

In Pokémon Red, Anthropic reports that the older Claude Sonnet 3.7 produced 413 files spread across 225 folders in about 10,000 steps. Battles and discovered items often received their own folder, while the evolution of the game rendered these archives obsolete. The model sometimes focused on organizing them rather than playing and only earned one badge.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Claude Fable 5, according to Anthropic, settled for a few files and earned four badges. It overlooked visible elements on the screen to record plans, strategies, and tested rules. However, the comparison remains limited to one game per model; all these trials on Claude come from Anthropic.

The company advises indicating to the model what its memory should be used for, then allowing it to decide how to utilize it. In a benchmark, it claims that strict prescriptions, such as retaining only errors, negated much of the advantages of more powerful models. Older versions, on the other hand, benefited from explicit instructions, such as reviewing their notes before responding.

Claude checks its notes during breaks

For Anthropic, the main risk arises when an agent believes a note is incorrect without verifying it. Dreaming, which builds on research from Letta and UC Berkeley, confronts notes with previous transcriptions during inactive moments. This review is intended to autonomously correct errors and organize knowledge into structured sets.

The trials in Pokémon provide concrete examples, according to the company: an inaccurate indication about the Pokémon Center led to repeated falls from a ledge. Another note placed a landmark on the wrong side of a mountain; Claude followed this indication three times, failed, and considered giving up. After Dreaming verified the transcription, the note was corrected, and the route was found again. Anthropic attributes the successful passage through Mount Moon to reach Celadon exclusively to this correction.

OpenAI also uses the name Dreaming

At OpenAI, Dreaming also refers to the memory of ChatGPT revamped in June. The system discreetly maintains a user profile and automatically updates outdated information. OpenAI claims a temporal accuracy improvement from 52.2% to 75.1%.

Researchers at Google and Google DeepMind use dreaming to work on an agent's search strategy: retained results serve to explore other solutions without re-querying the model. Google has also created WikiSkill, a wiki intended for agents. Like Anthropic, its organization distinguishes between original transcriptions, synthesized knowledge, and task-specific instructions.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.