Anthropic's recent research explores how its Claude language models can use idle-time "dreaming" to review, correct and reorganize previously recorded memories. Developers liken the process to human dreaming: selective offline consolidation could help models retain the most useful information and perform better on long-running, multi-session tasks.
The core problem
AI agents generate large numbers of logs and notes about their own activity. In many cases these records attempt to capture every detail, producing an unwieldy dataset in which important context can be lost, outdated information can persist, and incorrect facts become treated as established memory—degrading performance.
The "dreaming" approach
Anthropic's researchers propose that during idle periods agents should selectively revisit their memory files. In these offline reviews the system can filter out erroneous or redundant data, align notes with original sources, and build a structured, wiki-like knowledge base that is easier to consult in later sessions.
Pokémon Red tests: quantity of notes versus usefulness
The team tested the idea in Pokémon Red. They compared two models after roughly ten thousand in-game steps. Sonnet 3.7, introduced early last year, recorded almost everything: more than 400 files containing nearly one million characters, yet it managed to obtain only one badge in the game. By contrast, Claude Fable 5, released this year, wrote only ten files totalling about 40,000 characters, but collected four badges.
Anthropic researchers suggest Fable 5's advantage stemmed from documenting primarily navigation strategies, lessons learned and current goals, rather than storing information that could be directly read from the game interface.
Correcting mistaken memories: a concrete example
In another Pokémon trial a model misrecorded the location of a building and repeatedly became lost after leaving it. During an offline review the system compared its notes to the original log, detected the error, updated the memory file, and was then able to reach the next town using the corrected information.
Turning large document collections into a wiki
To test scalability, Anthropic ran the consolidation process on a large corporate corpus. They gave the model about 350,000 internal documents from the software company Stainless — roughly 420 million tokens. Over 128 sessions the model incrementally built and refined its own interconnected wiki.
Using that wiki improved the model's performance on a 25-question test from 13 percent to 48 percent.
In a separate experiment with corporate mergers-and-acquisitions documents that required integrating multiple sources, Claude answered questions with 87 percent accuracy when relying on its wiki, versus 73 percent accuracy using direct document search.
Costs and oversight
The background consolidation process makes frequently used facts directly accessible for future sessions while placing less-frequently needed details into searchable files. However, maintaining memory files consumes substantial compute resources, and excessive retained information can still overwhelm the system. Anthropic therefore emphasizes that automated memory maintenance requires careful oversight and engineering trade-offs.
Conclusions
The experiments indicate that more deliberate memory management during idle periods can reduce noise in agent notes, correct mistaken memories, and materially improve agents' ability to handle complex, multi-session tasks. At the same time, the computational and governance costs mean further work is needed before the technique can be broadly deployed.
An AI assistant contributed to the preparation of this article; the final content was edited and verified by a journalist.
