Anthropic research breaks down Claude agent memory, proposes "dreaming" mechanism
Anthropic published a research article titled "The components of agent memory," dividing Claude agent memory into four layers: session logs, memory notes, knowledge representations, and injected context, and introducing a "dreaming" mechanism that lets the model review logs, correct errors, and update its knowledge base between two interactions. In the CL-Bench database question-answering task, Claude Opus 5.5 and Sonnet 5.5 scored an average of 0.838 and 0.833 respectively with memory, compared with no-memory baselines of 0.655 and 0.644. In a test of about 350,000 enterprise documents, after scaling "dreaming" from 2 sessions to 128 sessions, the pass rate on 25 questions rose from 13% to 48%.
Why read this
AI score 60
Anthropic breaks agent memory into four layers and presents a "dreaming" mechanism, with specific figures from benchmarks and game tests, showing how the method can be put into practice.
Topics
Body · AI translation
Anthropic Research Says Claude AI Agents Can Get Smarter by "Dreaming"
2026-11-07 07:25 IT Home - Guyuan
Thanks to IT Home user Jingya Xianren for the tip submission!
IT Home, November 7 news: Anthropic yesterday (November 6) published a research article titled "The components of agent memory," systematically breaking down how Claude agents remember, filter, correct, and organize knowledge across sessions.
The research divides agent memory into four layers: session logs, memory notes, knowledge representations (such as a Wiki), and injected context, and introduces a "dreaming" mechanism that lets the model review logs, correct errors, distill rules, and update its knowledge base during idle time between two interactions.
The research team believes that AI can, like the human brain consolidating memories during sleep, gradually transform immediate experience into reusable long-term knowledge. Claude marks important events or writes memory notes during a session, but not all information is worth saving; it needs to judge which content will truly be useful in the future.
Anthropic points out that memory is essentially a prediction about the future; for example, recording a verified rule means the model expects to rely on it again in the future.
In the database question-answering task of the continual learning benchmark CL-Bench, Claude models with memory capabilities performed significantly better than no-memory baselines. After using memory, Claude Opus 5.5 and Sonnet 5.5 achieved average scores of 0.838 and 0.833, respectively, while their no-memory baselines were 0.655 and 0.644. Stronger models can autonomously form a learning loop of "failure - reflection - verification - rule summarization."
In a Pokemon Red game test, Anthropic also found that "less but better" memory is more effective. Claude Fable 5 maintained only 10 memory files yet earned 4 badges, while Sonnet 3.7 wrote 413 files and earned only 1 badge. The research concludes that models should prioritize saving strategies, plans, and tested rules rather than constantly changing screen states.
The "dreaming" mechanism also demonstrated error-correction capabilities. When Claude incorrectly recorded the location of a Pokemon Center and got lost as a result, the system corrected the memory by comparing it against the original session records, allowing it to eventually reach the next city successfully; a run without memory correction repeatedly doubled back and failed to complete the objective.
In a test with about 350,000 enterprise documents, Claude gradually 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 says agent memory still faces challenges such as when to write, how to keep it updated, and how to avoid error accumulation. In the future, AI may become better at "reviewing experience" after completing tasks, transforming scattered interactions into transferable, verifiable long-term knowledge.
IT Home attaches the reference address
The components of agent memory
Related reading:
Advertising statement: External redirect links contained in the article (including but not limited to hyperlinks, QR codes, passwords, and other forms) are used to convey more information and save selection time; the results are for reference only. All articles on IT Home containing external links include this statement.