Anthropic: selective memory and the Dreaming process improve AI agent performance across sessions
The company Anthropic has published research on AI agent memory: selectively writing notes works better than archiving everything. In a test with Pokémon Red, the Claude Fable 5 model with selective memory obtained 4 badges, while the Claude Sonnet 3.7 model with a complete archive got only 1.
The company Anthropic published research on how the Claude model retains knowledge between sessions through text notes saved to a folder. According to the company, the key problem is selection - the agent must already estimate during the task what will be useful later. In a test with the game Pokémon Red, the older Claude Sonnet 3.7 model, which recorded practically everything, created 413 files in 225 folders over approximately 10 000 steps of the game; the archive quickly became outdated, and the model spent part of its time just reorganizing notes instead of playing. It finished the game with one badge. The newer Claude Fable 5 model kept only a few files with plans, strategies, and verified rules, and did not record information visible directly on the screen; it obtained four badges. Anthropic notes that the comparison is based on only one playthrough of the game per model.
According to Anthropic, it works better when the model is explained the purpose of memory and given freedom in deciding how to use notes. In one benchmark, strict instructions such as \"record only errors\" eliminated most of the advantage of selective memory for stronger models, while for older models, clear instructions - for example, an instruction to read notes before every response - helped instead.
The second part of the research is a process called Dreaming, which builds on research by the companies Letta and UC Berkeley. During a break between tasks, the Claude model goes through transcripts of earlier sessions and compares them with its notes in order to correct errors. In the Pokémon Red test, the model had incorrectly noted which side of a mountain a certain point was located on, causing it to fall into the same trap three times; after going through the Dreaming process, it detected the error and found the correct path. According to the company, Dreaming can also organize larger knowledge bases - from approximately 350 000 internal documents of the company Stainless, which Anthropic acquired, the model created an interconnected wiki. In a separate legal test by Anthropic, the model using this wiki correctly answered 87 percent of questions requiring information from multiple documents, while with plain retrieval it was 73 percent; for simple retrieval tasks, the two approaches were comparable.
At the same time, Anthropic points to open problems: writing notes during a task distracts the agent, the Dreaming process costs additional computing power and its results only show up with a delay, and both notes and the wiki require regular maintenance, otherwise errors and hallucinations accumulate. The company also warns against automatically loading too many memories into every session, since this takes up context and carries any erroneous notes over into all future tasks.
Why it matters
The research offers concrete guidance for designing AI agents that function well over the long term: instead of maximizing the volume of stored data, it pays off to select only strategies and rules and let stronger models decide for themselves what to remember. The Dreaming process also addresses the risk of an agent blindly trusting an error once written down and repeating it in every subsequent session - this is relevant for anyone deploying agents on multi-year or multi-day projects.
Two audiences, two different impacts
What this means
For individuals
When working with AI agents on longer projects, it pays off to design notes selectively (strategies, rules, exceptions) instead of logging every step, and for stronger models, to give them freedom to decide what to write down.
For a business
Companies deploying AI agents on multi-day tasks can reduce the risk of error accumulation in memory and unnecessary costs from reorganizing data by introducing periodic review of notes, for example through a process like Dreaming, which is already available as a research preview for Claude Managed Agents.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.