正文 · AI 翻译
Thanks to IT Home reader 惊讶仙人 for the tip!
IT Home reported on November 7 that Anthropic published a research article yesterday (November 6), “The components of agent memory,” systematically breaking down how Claude agents remember, select, correct, and organize knowledge across sessions.
The research says AI agents can draw on the human brain’s mechanism of “organizing memories during sleep,” reflecting on past records, correcting errors, and building knowledge structures during idle time, thereby significantly improving performance on long term tasks.
The research divides agent memory into 4 layers: session logs, memory notes, knowledge representations (such as a Wiki), and injected context. It also introduces a “dreaming” mechanism that lets the model review logs, correct errors, distill rules, and update its knowledge base during idle time between interactions.
The research team believes AI can gradually turn immediate experiences into reusable long term knowledge, much as the human brain consolidates memories during sleep. Claude marks important events or writes memory notes during a session, but not all information is worth saving; it needs to judge what will actually be useful in the future.
Anthropic says memory is essentially a prediction about the future: recording a verified rule, for example, means the model expects to rely on it again.
In the database question answering task of the continual learning benchmark CL-Bench, Claude models with memory performed markedly better than baselines without memory. With memory, Claude Opus 5.5 and Sonnet 5.5 achieved average scores of 0.838 and 0.833, respectively, compared with their respective baselines without memory of 0.655 and 0.644. Stronger models can independently form a learning loop of “failure — reflection — verification — rule summarization.”
In tests using the game Pokémon Red, Anthropic also found that a smaller, more carefully selected set of memories was more effective. Claude Fable 5 maintained only 10 memory files yet earned 4 badges; Sonnet 3.7 wrote 413 files but earned only 1 badge. The research suggests that models should prioritize saving strategies, plans, and tested rules over constantly changing screen states.
The “dreaming” mechanism also demonstrated an ability to correct errors. When Claude incorrectly recorded a Pokémon Center’s location and got lost as a result, the system corrected the memory by comparing it with the original session record, allowing Claude to eventually reach the next city; the run with uncorrected memory repeatedly doubled back and failed to reach its goal.
In a test involving about 350,000 corporate documents, Claude gradually built a Wiki through multiple rounds of “dreaming.” As the scale of “dreaming” grew from 2 sessions to 128, the pass rate on 25 questions rose from 13% to 48%; combining the Wiki with a search tool raised the overall score from 54% to 61%.
Anthropic says agent memory still faces challenges, including when to write information, how to keep it up to date, and how to prevent errors from accumulating. In the future, AI may become better at “reviewing experiences” after completing tasks, turning scattered interactions into transferable, verifiable long term knowledge.
IT Home provides the reference link below
The components of agent memory
Related reading:
“Goodbye, knowledge cutoff date: Google teaches AI to ‘sleep,’ learn as it goes, and retain old knowledge”
Advertising disclosure: External links in this article (including but not limited to hyperlinks, QR codes, and passcodes) are provided to convey more information and save time in evaluating options. The results are for reference only. All IT Home articles containing external links include this disclosure.