When an AI agent returns to a project, its notes can help it pick up where it left off. But saving everything can create a different problem: outdated or mistaken information may crowd out what is useful. Anthropic’s tests of Claude point to a more selective approach, with a later review process called Dreaming intended to check and improve what the agent remembers.
Memory works best when it is selective
Anthropic explored how Claude could keep information in files between sessions without changing the model itself. The challenge is deciding what to write down before knowing what will matter later. A note is a guess about what may be useful in a future task.
A Pokémon Red test showed the cost of treating every detail as worth saving. Claude Sonnet 3.7 created 413 files in 225 folders over roughly 10,000 game steps, often making separate files for individual battles and items. Since the game state kept changing, much of the archive quickly became outdated. The model sometimes spent its effort reorganizing notes instead of playing, and finished with just a single badge.
Claude Fable 5 used a smaller set of files and earned four badges. It skipped information already visible on screen and recorded plans, strategies, and rules it had tested. Anthropic cautions that this comparison covers only a single playthrough per model, so it is a limited basis for comparing their performance.
Give agents room to decide what matters
Anthropic recommends explaining why an agent has memory and allowing it to choose how to use it. In one benchmark, rigid directions such as “only record mistakes” removed much of the advantage for stronger models. Older models, by contrast, benefited from direct instructions, including a reminder to read their notes before every response.
The distinction matters because a memory rule can shape what the agent notices and preserves. A strategy that works for one model may constrain another. Anthropic’s findings suggest that memory instructions should account for the model’s capabilities rather than assume one fixed rule will suit every agent.
Dreaming checks notes against experience
Even carefully selected notes can contain errors. If an agent trusts a false memory in a later session, it may repeat a mistake. Anthropic’s Dreaming process is designed to review transcripts from earlier sessions during an idle period and compare those records with the notes.
In the Pokémon test, Claude had recorded the wrong side of a mountain for a landmark. It relied on that note, fell into the same trap three times, and was close to giving up before reaching its goal. After a Dreaming review, Claude checked the note against the transcript, caught the error, and found the right path.
Dreaming can also help organize larger collections. Anthropic says Claude built a linked wiki from about 350,000 internal documents from Stainless, a software company Anthropic acquired. The process is intended to create pages while addressing gaps, duplicates, and broken links. Anthropic reported the greatest benefit when Claude could use the wiki alongside search, making the organized pages a complement to search.
In a separate legal test by Anthropic, the wiki was most helpful for questions whose answers were spread across multiple documents: Claude scored 87 percent with the wiki, compared to 73 percent with search alone. For simple lookups, the wiki performed roughly on par with search, and in one case it grew larger than the source documents.
Useful memory still needs upkeep
Anthropic describes several costs and open questions. Taking notes during a task can distract an agent. Dreaming uses extra compute, and its improvements arrive after the review rather than immediately. Notes and wikis also need maintenance, since errors and hallucinations can build up over time.
Loading too many memories into every new session creates another burden: notes use context and can carry a bad entry into future work. For agent designers, the practical lesson from Anthropic’s tests is to keep information that supports future decisions, review it against past experience, and avoid making every memory part of every session.
Anthropic introduced Dreaming in May as a research preview for Claude Managed Agents, its cloud platform for AI agents. The service reads an existing memory store and up to 100 past sessions, removes duplicate and outdated entries, then creates a reorganized store while leaving the original untouched. At launch, Dreaming supported only Claude Opus 4.7 and Sonnet 4.6.