Claude Agents Fix Bad Memories by 'Dreaming'

Anthropic on Nov. 6 published research showing how its Claude models decide what to remember between sessions, and how an offline review step it calls "dreaming" corrects bad notes and builds knowledge wikis. In one test, as Claude agents scaled a wiki from two to 128 dreaming sessions, scores on 25 questions about a company's internal documents rose from 13% to 48% of graded criteria.

Key takeaways

• On the CL-Bench 1.0 database task, memory raised Claude Opus 5.5's mean reward from 0.655 to 0.838 and Claude Opus 4.7's from 0.383 to 0.700.
• Strict rules such as "record only failures" erased much of the memory gain for Opus 5 and Opus 4.7; Anthropic says newer models do better when left to judge what matters.
• Dreaming, a research-preview feature in Claude Managed Agents, re-checks notes against session transcripts; it fixed a navigation error in a Pokémon run, and revised wiki instructions lifted a contract brief from 72% to 100% of graded facts.

How Claude agent memory works

Anthropic defines agent memory as knowledge Claude gains from experience and keeps between sessions as files, not as changes to model weights. Its hierarchy runs from raw transcripts at the bottom, to notes Claude writes in-session, to wikis, to a small set of notes loaded into every new session. Dreaming, an idea the post links to "sleep-time compute" research from Letta and UC Berkeley, uses idle time to reorganize those layers.

CL-Bench results: memory gains by model

Anthropic ran the CL-Bench 1.0 database task on Claude Managed Agents. Each question began in a fresh session, and the only carryover was a memory folder. After each answer, Claude saw the correct solution and could write notes.

Anthropic said Opus 5 was the first model to "close the loop," writing rules, the evidence behind them and warnings about traps. Older models benefited most from steering: prompting Claude Sonnet 4.6 to read its memory and test its theories took it from 0.330 to 0.641, versus 0.364 with memory alone.

"We've found that it helps to tell Claude what memory is for, but not how to use it," the researchers wrote.

What Pokémon showed about salience

About 10,000 steps into Pokémon Red, Claude Sonnet 3.7 had written 413 memory files across 225 folders and earned one badge. Claude Fable 5 kept nine files in two folders plus a single global.md file and earned four badges.

In another run, Claude wrote down the wrong side of the mountain for a landmark near Mt. Moon and repeatedly fell off a one-way ledge. After dreaming checked the note against the transcript and rewrote it, Claude went back into the cave and reached Cerulean City. The uncorrected run never did.

Wikis, search and a 20,000-token saving

For the wiki test, Anthropic used about 350,000 internal files, roughly 420 million tokens, from Stainless, a software company it acquired in May. Combining the wiki with search raised scores from 54% to 61% as the wiki grew, against 53% for search alone, and saved about 20,000 tokens per question.

The gains were uneven. On 24 merger questions, multi-document items scored 87% with the wiki versus 73% with search, but on one legal-brief task the wiki showed no clear gain.

Context: dreaming in Claude Managed Agents

Anthropic introduced dreaming for Claude Managed Agents in May at its Code with Claude conference, SiliconANGLE reported. According to the developer documentation, a dream takes an existing memory store and 1 to 100 past sessions, writes a new store without touching the original, and can take from minutes to a few hours.

Dreams are billed at standard token rates for the chosen model and remain in research preview, with access by request. The research post lands weeks after Claude Opus 5.5 arrived 20% cheaper.

What's next

Anthropic listed open questions, including when agents should write memory and how often dreaming should run given its compute cost. "No model can really know what will happen in future turns," the researchers wrote. The company has not given a date for dreaming to leave research preview.

FAQ

What is dreaming in Claude?

Dreaming is an offline job in which Claude reviews past session transcripts and an existing memory store, then produces a reorganized store with duplicates merged, errors corrected and new notes or wikis added.

Can developers use dreaming now?

Only through a research preview of Claude Managed Agents on the Claude API, which requires an access request. It is billed at standard token rates.

Should agents be told exactly what to remember?

Anthropic's tests suggest not for newer models. A fixed "record only failures" rule cut much of the memory benefit for Opus 5 and Opus 4.7, while older models such as Sonnet 4.6 gained from explicit prompting.

References

• The components of agent memory — Anthropic (Nov. 6, 2026)
• Dreams — Claude Platform Docs
• Anthropic is letting Claude agents 'dream' so they don't sleep on the job — SiliconANGLE (May 6, 2026)