- Published on
Anthropic Study Says Claude AI Agents Can Become Smarter by 'Dreaming'
- Authors
- Name
- DeepHorus
On November 6, Anthropic published a research article titled "The components of agent memory." The article systematically breaks down how Claude agents remember across sessions, filter, correct errors, and organize knowledge. The research argues that agents can draw on the human brain's mechanism of "sleep-based memory consolidation," reflecting on historical records during idle time and updating their knowledge bases.
The team calls this process "dreaming." It lets the model review logs, correct errors, distill rules, and update its knowledge base between interactions. The research says this is expected to improve agent performance on long-horizon tasks.
The research divides agent memory into four layers: session logs, memory notes, knowledge representations (such as a Wiki), and injected context. Claude marks important events or writes memory notes, but it needs to judge which information will truly be useful in the future. The team points out that memory is essentially a prediction about the future.
In the CL-Bench database question-answering task, Claude with memory clearly outperformed the no-memory baseline. Claude Opus 5.5 and Sonnet 5.5 achieved average scores of 0.838 and 0.833, respectively, when using memory, compared with 0.655 and 0.644 for the no-memory baseline. Stronger models can autonomously form a closed loop of "failureâreflectionâverificationâsummarizing rules."
In the "Pokémon Red" test, "less but better" memory was more effective. Claude Fable 5 maintained only 10 memory files and earned 4 badges; Sonnet 3.7 wrote 413 files and earned only 1 badge. The research argues that models should prioritize saving strategies, plans, and tested rules rather than screen states.
The "dreaming" mechanism also demonstrated error-correction capabilities. Claude once incorrectly recorded the location of a Pokémon Center and got lost. After the system compared the original session records and corrected the memory, it reached the next city. A run without corrected memory repeatedly doubled back and failed to complete the objective.
In tests with approximately 350,000 enterprise documents, Claude built a Wiki through multiple rounds of "dreaming." As the scale of "dreaming" expanded from 2 sessions to 128 sessions, the pass rate on 25 questions rose from 13% to 48%. After combining the Wiki with search tools, the overall score rose from 54% to 61%.
Anthropic said agent memory still faces challenges such as when to write, how to update, and how to avoid the accumulation of errors. In the future, AI may become better at "reviewing experience" after completing tasks, turning scattered interactions into transferable, verifiable long-term knowledge. This direction provides new ideas for continual learning in agents.
