# Anthropic on the components of Claude's agent memory [Canonical reading](https://escapemeridian.com/readings/anthropic-on-the-components-of-claude-s-agent-memory-6df79aa7) Anthropic describes agent memory as knowledge Claude gains through experience and retains between sessions, distinct from changes to model weights or knowledge gained within one conversation. Its central idea is that agents can record experience during a task, then use idle time—called “dreaming”—to review logs, correct notes, and build or maintain memory representations such as wikis. In Anthropic’s examples, memory helped models learn rules across separate database questions and retain strategies while playing Pokémon. Saving more was not always better: one model’s large collection included details that went out of date and consumed time without advancing the game. A smaller, more selective memory supported further progress. In document tasks, wikis helped most when answers required combining information across documents; they offered no clear gain when a single document sufficed, and one wiki grew larger than its source files. Anthropic also describes trade-offs: memory can be wrong or become outdated, reviewing it takes compute, and automatically loading notes into every session uses context and can carry errors forward. The authors leave open when agents should write memories and how often they should review them. These evaluations illustrate possible approaches, but do not establish how well they will work across other tasks or changing environments. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source Anthropic · Article Original source date: Nov 6, 2026 (2026-11-06T00:00:00.000Z). Date basis: Date stated by the publisher. Added to Escape Meridian: 2026-11-09T04:33:59.000Z. First observed: 2026-11-06T23:43:07.325Z. Retrieved: 2026-11-09T04:33:01.496Z. - [The components of agent memory](https://www.anthropic.com/research/components-of-agent-memory) ## Limits The reported evaluations cover specific benchmark, game, and document tasks; the post does not establish that the results generalize to other tasks or environments. The article raises unresolved questions about when to create memory and how to keep it up to date; these evaluations do not settle those design choices. ## Open question How should Claude decide when to write memory during a task versus defer memory creation to dreaming, given the trade-off between immediate access and distraction? ## Source-stated Anthropic describes agent memory as knowledge Claude gains from experience and retains between sessions, and distinguishes it from other forms of model learning or context management. > By “agent memory,” we mean knowledge that Claude gains from experience—including interactions with users and tasks it’s been assigned—and that it retains between sessions - [Supporting source](https://www.anthropic.com/research/components-of-agent-memory) ## Source-stated The post describes dreaming as using idle time to review logs and organize existing context, including correcting mistakes or producing notes and representations. > models to use idle time between interactions to reorganize their existing context. The same idea can help agents organize their memory by reflecting on logs - [Supporting source](https://www.anthropic.com/research/components-of-agent-memory) ## Source-stated In the Pokémon example, Sonnet 3.7 created many memory files, including details that went out of date, while Fable 5 used a smaller set focused on plans and tested rules and progressed further. > With a much smaller memory folder, Fable 5 progressed much further in the game than Sonnet 3.7. - [Supporting source](https://www.anthropic.com/research/components-of-agent-memory) ## Source-stated The authors report that wikis helped more on questions requiring information across several documents, while one task showed no clear gain over search. > Claude scored 87% when it consulted the wiki, compared with 73% when it used search alone. - [Supporting source](https://www.anthropic.com/research/components-of-agent-memory) ## Source-stated The authors identify unresolved choices about when to create memory and how to keep it up to date, noting that writing during a task can divert attention and dreaming may not make results immediately accessible. > logging notes in the moment inevitably draws Claude’s attention away from its main task. In contrast, dreaming can push memory creation to the background - [Supporting source](https://www.anthropic.com/research/components-of-agent-memory) ## Source and contributor context Anthropic · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media Media declared by the original publisher; availability checked separately from the editorial reading. - [Illustration of two offset half circles with a grid pattern, one green and one pink, facing each other across a gap on a light blue background. · Anthropic](https://www.anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F31b7b06af6c391aa4f93ad9747d00ae248ea3c63-2000x1125.jpg&w=1920&q=75) ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Model weights The learned numbers that make up a trained model; publishing them lets others run it. - [Definition reference](https://developers.google.com/machine-learning/glossary#weight) ## Benchmark A defined set of tests used to compare systems. Results depend on the tasks and measures selected. - [Definition reference](https://arxiv.org/abs/2211.09110) ## Anthropic Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems. - [Explore this context](/entities/anthropic) - [Reference](https://www.anthropic.com/news) - [Reference](https://www.anthropic.com/research/team/economics) ## Claude Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding. - [Explore this context](/entities/claude) - [Reference](https://claude.com/product/overview) - [Reference](https://www.anthropic.com/news) - [Reference](https://www.anthropic.com/research/team/economics) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 01689bb6-18a4-4fd8-823d-ab279055a59d. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-09T04:33:59.000Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 43d86424f0e08ba69cc2b4acf09366505bd2f4c4414a5b111684493a57689e70. Source text hash: a5c58a790b980616e143f95782fc4032813acd385fc6ce1ec334bf830d7dc20c. ## Revision and source-check history ### Version 1 — 2026-11-09T04:33:59.000Z Anthropic describes agent memory as knowledge Claude gains through experience and retains between sessions, distinct from changes to model weights or knowledge gained within one conversation. Its central idea is that agents can record experience during a task, then use idle time—called “dreaming”—to review logs, correct notes, and build or maintain memory representations such as wikis. In Anthropic’s examples, memory helped models learn rules across separate database questions and retain strategies while playing Pokémon. Saving more was not always better: one model’s large collection included details that went out of date and consumed time without advancing the game. A smaller, more selective memory supported further progress. In document tasks, wikis helped most when answers required combining information across documents; they offered no clear gain when a single document sufficed, and one wiki grew larger than its source files. Anthropic also describes trade-offs: memory can be wrong or become outdated, reviewing it takes compute, and automatically loading notes into every session uses context and can carry errors forward. The authors leave open when agents should write memories and how often they should review them. These evaluations illustrate possible approaches, but do not establish how well they will work across other tasks or changing environments. ## Original source Anthropic · Article Original source date: Nov 6, 2026 (2026-11-06T00:00:00.000Z). Date basis: Date stated by the publisher. Added to Escape Meridian: 2026-11-09T04:33:59.000Z. First observed: 2026-11-06T23:43:07.325Z. Retrieved: 2026-11-09T04:33:01.496Z. - [The components of agent memory](https://www.anthropic.com/research/components-of-agent-memory) ## Limits The reported evaluations cover specific benchmark, game, and document tasks; the post does not establish that the results generalize to other tasks or environments. The article raises unresolved questions about when to create memory and how to keep it up to date; these evaluations do not settle those design choices. ## Open question How should Claude decide when to write memory during a task versus defer memory creation to dreaming, given the trade-off between immediate access and distraction? ## Source-stated Anthropic describes agent memory as knowledge Claude gains from experience and retains between sessions, and distinguishes it from other forms of model learning or context management. > By “agent memory,” we mean knowledge that Claude gains from experience—including interactions with users and tasks it’s been assigned—and that it retains between sessions - [Supporting source](https://www.anthropic.com/research/components-of-agent-memory) ## Source-stated The post describes dreaming as using idle time to review logs and organize existing context, including correcting mistakes or producing notes and representations. > models to use idle time between interactions to reorganize their existing context. The same idea can help agents organize their memory by reflecting on logs - [Supporting source](https://www.anthropic.com/research/components-of-agent-memory) ## Source-stated In the Pokémon example, Sonnet 3.7 created many memory files, including details that went out of date, while Fable 5 used a smaller set focused on plans and tested rules and progressed further. > With a much smaller memory folder, Fable 5 progressed much further in the game than Sonnet 3.7. - [Supporting source](https://www.anthropic.com/research/components-of-agent-memory) ## Source-stated The authors report that wikis helped more on questions requiring information across several documents, while one task showed no clear gain over search. > Claude scored 87% when it consulted the wiki, compared with 73% when it used search alone. - [Supporting source](https://www.anthropic.com/research/components-of-agent-memory) ## Source-stated The authors identify unresolved choices about when to create memory and how to keep it up to date, noting that writing during a task can divert attention and dreaming may not make results immediately accessible. > logging notes in the moment inevitably draws Claude’s attention away from its main task. In contrast, dreaming can push memory creation to the background - [Supporting source](https://www.anthropic.com/research/components-of-agent-memory) ## Source and contributor context Anthropic · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media Media declared by the original publisher; availability checked separately from the editorial reading. - [Illustration of two offset half circles with a grid pattern, one green and one pink, facing each other across a gap on a light blue background. · Anthropic](https://www.anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F31b7b06af6c391aa4f93ad9747d00ae248ea3c63-2000x1125.jpg&w=1920&q=75) ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Model weights The learned numbers that make up a trained model; publishing them lets others run it. - [Definition reference](https://developers.google.com/machine-learning/glossary#weight) ## Benchmark A defined set of tests used to compare systems. Results depend on the tasks and measures selected. - [Definition reference](https://arxiv.org/abs/2211.09110) ## Anthropic Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems. - [Explore this context](/entities/anthropic) - [Reference](https://www.anthropic.com/news) - [Reference](https://www.anthropic.com/research/team/economics) ## Claude Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding. - [Explore this context](/entities/claude) - [Reference](https://claude.com/product/overview) - [Reference](https://www.anthropic.com/news) - [Reference](https://www.anthropic.com/research/team/economics) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 01689bb6-18a4-4fd8-823d-ab279055a59d. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-09T04:33:59.000Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 43d86424f0e08ba69cc2b4acf09366505bd2f4c4414a5b111684493a57689e70. Source text hash: a5c58a790b980616e143f95782fc4032813acd385fc6ce1ec334bf830d7dc20c. ### passed — 2026-11-09T04:33:59.000Z — anthropic/claude-opus-5.5 Publication: 01689bb6-18a4-4fd8-823d-ab279055a59d attribution — pass: The draft attributes the framing, evaluations and trade-offs to Anthropic and its authors. It does not present their results as independent facts. The publisher page is Anthropic's research post. > We’ve found that dreaming has been a useful tool for helping us give Claude effective memory. summary — pass: The draft leads with Anthropic's definition of agent memory and its central idea of dreaming, and explains both in plain language. It accurately covers the database task, the Pokémon contrast, where wikis helped or did not, and the stated trade-offs. The question follows from the source's open issue about when to write memory. Omitting the Stainless scaling results and the steering findings does not distort the source's meaning. > Agent memory can use a similar division of labor. Agents capture much of their work in logs or transcripts. date — pass: evidence.originalPublished is 2026-11-06, based on the publisher page. It matches the feed date, and nothing in the evidence indicates an earlier first publication. support — pass: Each claim's excerpt is an exact source quote that supports it. The source supports the stated trade-offs: injected notes use up tokens and can carry errors into every session, and dreaming requires compute. The limitations are framed as gaps rather than invented facts, and the question is not leading. > Injected memories use up valuable tokens, and an incorrect note can be carried into every session. media — pass: The selected media is the verified direct link to the original Anthropic article, which is the most direct way to access the source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/6df79aa741416fe2b361bdbea29cf0e6803c4bcb04141d5121f2b6262072ec58/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/anthropic-on-the-components-of-claude-s-agent-memory-6df79aa7.json)