🔍 Read the full analysis: How To Understand AI Memory In Agents: Research And Practical Lessons on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Anthropic published a research post on November 6, 2026, describing how Claude agents can save, organize and revise information between sessions. Its examples suggest that concise, structured notes and checks against earlier records can help on some tasks, but the results are company evaluations, several based on single runs, and do not establish a general performance guarantee.
Anthropic published a research post on November 6, 2026, outlining how its Claude models can save information between conversations, retrieve it later and revise it when it proves unreliable. The post describes a memory system for AI agents—not a change to the models’ trained weights—and presents company-run tests in which memory structure and selection were associated with different results.
Anthropic calls agent memory knowledge a model gains from interactions or assigned tasks and retains between sessions in a form it can read later. It distinguishes that stored information from a model’s context window, which contains material available during the current session, and from summaries used to continue work after a context window fills. The post also excludes changes to model weights during training.
The post describes several components: memory notes that record useful information; salience judgments about what is likely to matter again; a between-session review process Anthropic calls “dreaming”; and structures such as linked wiki pages. A hierarchy can keep a short overview available automatically while leaving detailed notes to be retrieved only when needed. Anthropic relates this division of work to sleep-time compute research proposed by researchers at Letta and the University of California, Berkeley.
In one Pokémon Red comparison, Anthropic reports that Claude Sonnet 3.7 created 413 files across 225 folders and had earned one badge at about 10,000 game steps. Claude Fable 5 created nine files in two folders plus a top-level dashboard and had earned four badges. Anthropic says each model was represented by one run. The company also reports that a self-built wiki increased performance on 25 questions about company documents from 13% to 48% of graded criteria over 128 sessions. These are Anthropic’s evaluations, published in a research post rather than a peer-reviewed paper.
Infographic 01 · Remember less, remember the right things
What the screen shows does not belong in memory
Money and position can be read off the screen every turn. Plans, strategy and tested rules exist only if the agent writes them down.
Claude Fable 5 wrote nine files in two folders, plus one top-level dashboard file (global.md) with its goal, plan and rules for using memory.
Memory files at about 10,000 steps
One run each, Pokémon Red, a memory folder the model could read and write
Over 17 steps, Sonnet 3.7 created six files, deleted nine others and took no actions in the game.
Source: Anthropic, The components of agent memory (research post, 6 Nov 2026). One run per model, Anthropic’s own evaluation.
Powered by Thorsten Meyer AI
Memory Design Can Shape Agent Results
The findings address a practical problem for anyone using an agent across multiple sessions: the system must decide what to retain, how to locate it, and when to revise it. A larger archive is not automatically more useful. In Anthropic’s game example, the newer model’s smaller collection was accompanied by more progress, while the older model spent some steps creating and deleting files without acting in the game. That comparison is suggestive, but one run per model cannot establish that a smaller memory will consistently perform better.
Better memory could make agents more useful for recurring work, such as following project rules or answering questions about a document collection. It also creates risks. An incorrect note may steer later answers; an outdated instruction can survive after circumstances change; and an agent can confuse information from different tasks. The practical value of memory depends not just on storage, but on accuracy, retrieval and correction.
The post’s results give developers a set of design questions rather than a universal recipe: what information is genuinely new, what can be retrieved from the current task, and how can a saved claim be checked? Anthropic’s examples support testing these choices on a specific workflow, rather than assuming the reported scores will carry over to other models or uses.
Products Worth Considering
Apple MacBook Air 15-inch (2025) with M4 Chip, 24GB Memory, 512GB SSD, Midnight
Apple iMac (2024) 24-inch All-in-One Desktop with M4 Chip, 16GB Unified Memory, 256GB SSD, Green
Apple iMac (24-inch, 2024) with M4 Chip, 16GB Unified Memory, 256GB SSD, Silver
As an affiliate, we earn on qualifying purchases.
From Session Context to Saved Notes
In a conventional chat, the model responds using the material available in its current context window. When that session ends, information is not necessarily carried into the next conversation. Agent memory adds a persistent store that the model can consult later, making the system’s own choices about what to save part of its behavior.
Anthropic’s post frames the process as more than writing notes. An agent may record important events as they happen, review transcripts between sessions, correct or remove mistaken entries, and arrange retained knowledge into searchable representations. The proposed “dreaming” step draws on the idea that useful learning can happen outside the immediate task. The article’s practical lessons are separately attributed to the publisher’s own experience running a small group of Claude-based agents; Anthropic did not review or sponsor that account.
““knowledge that Claude gains from experience—including interactions with users and tasks it’s been assigned—and that it retains between sessions, saved as context it can read later.””
— Anthropic, defining agent memory in its November 6 research post
How Broadly the Results Apply
The reported evaluations do not establish how reliably these methods work across tasks, models or real-world deployments. Anthropic says several results are based on single runs, including the Pokémon comparison; the post is not a peer-reviewed paper. The document-question result is reported as a change in the share of graded criteria met, not as a general accuracy measure for all retrieval tasks.
It is also unclear from the supplied account how the systems would perform over longer periods as their stores grow, or how consistently they identify and remove sensitive, stale or conflicting information. The publisher’s own agent lessons—such as index limits, stale notes and lost detail in handoffs—are practitioner observations, not controlled findings attributed to Anthropic. Broader comparisons and independent replication would be needed to establish which memory designs work best.
More Tests Needed Across Tasks
Anthropic’s post presents a framework and evaluation examples, rather than announcing a product release or a fixed standard for agent memory. The next useful evidence would come from repeatable comparisons across more tasks and runs, including tests of whether agents can detect incorrect or outdated notes and preserve useful information without accumulating unnecessary files.
For users and developers, the immediate practical step is to treat saved memory as fallible: inspect what an agent retains, date rules that may change, and check important answers against their original sources. The source material does not give a schedule for further Anthropic evaluations, so it remains unclear when additional results or independent studies may appear.
Key Questions
What does Anthropic mean by AI agent memory?
It means information an agent gains from tasks or interactions, saves between sessions and can read later. Anthropic distinguishes it from information available only in the current context window.
What is “dreaming” in the research post?
Anthropic’s term for an agent reviewing logs or transcripts between sessions to correct, organize or add to its notes. It is a described system process, not a claim that the model sleeps.
Did the Pokémon test prove that smaller memory stores are better?
No. Anthropic reports one run each for the two models in that comparison. The results show what happened in those runs, but do not establish a general rule that smaller memory always performs better.
Were the reported findings independently verified?
The supplied source describes the results as Anthropic’s own evaluations, published in a research post rather than a peer-reviewed paper. It does not report independent replication.
Source: ThorstenMeyerAI.com
