Read the full analysis: AI Agents With Memory: What The Research Shows In Practice on ThorstenMeyerAI.com
TL;DR
Anthropic published a research post on Nov. 6 describing components of AI agent memory, including saved notes, retrieval structures and a process for correcting notes against past transcripts. Its reported evaluations suggest that compact, organized memory can outperform a much larger collection of files, but the results are company-reported and include single-run comparisons.
Anthropic published research on Nov. 6 describing how its Claude models can retain and use information between sessions, with examples suggesting that a smaller, organized memory store can work better than an unstructured mass of notes. The post, titled The components of agent memory, reports results from company-run tests; it is a research post, not a peer-reviewed paper.
Anthropic distinguishes an agent’s working context from its longer-term memory. The context window contains what a model can use in the current session; agent memory is information saved between sessions and made available later. The company’s definition excludes model-weight changes, learning confined to one conversation and summaries used to continue within a new context window.
The post describes several parts of a memory system: memory notes that preserve potentially useful information; salience judgments about what is worth saving; a between-session review process Anthropic calls “dreaming”; and organized representations, such as linked wiki pages. A small, frequently loaded top layer can point the agent toward other material, which it searches or reads only when needed.
In a Pokémon Red test, Anthropic compared one run each of Claude Sonnet 3.7 and Claude Fable 5 at about 10,000 steps. Sonnet 3.7 had created 413 files across 225 folders and earned one badge. Fable 5 had created nine files in two folders plus a top-level dashboard and earned four badges. Anthropic also reports that a self-built wiki improved performance on 25 questions about company documents, from 13% to 48% of graded criteria over 128 sessions.
Agent memory · game test
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
Claude Sonnet 3.7225 folders · 1 badge413 filesClaude Fable 52 folders + global.md · 4 badges9 files
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
Why Memory Design Changes Results
The findings address a practical weakness in assistants used repeatedly: a model may have access to a large amount of saved information yet fail to retrieve the right detail, or rely on a note that is wrong or out of date. Anthropic’s examples point to selection and organization, not simply storage capacity, as central design problems.
That matters for users and organizations deploying agents across multi-session work. A system that records a durable plan, tested rule or user preference may be more useful than one that saves every exchange. But the results do not establish that memory will reliably improve every task. They are company-reported evaluations, and the Pokémon comparison involved one run per model, limiting what can be inferred from the badge difference.
Memory also creates a risk beyond forgetting: an agent can preserve an incorrect inference and apply it later as if it were established. Anthropic’s proposed review process attempts to address that by checking saved notes against original logs. The post presents this as a system component, not proof that errors are always found or corrected.
From Session Context to Saved Notes
The research post builds on a distinction familiar to users of chat systems. Information in a conversation is available while it remains in the model’s context, but a new session may begin without it unless a separate memory mechanism carries details forward. Persistent agent memory stores selected information outside that immediate exchange and brings it back into later work.
Anthropic borrows the term “dreaming” for an offline review stage that can revisit transcripts, correct notes and create more useful representations. The post connects this approach to the idea of sleep-time compute, which researchers at Letta and UC Berkeley have proposed. The analogy is a design framing; it does not mean an AI agent sleeps or has human memory.
Examples in the post also stress the difference between information visible at every step and knowledge that must be learned. In the game test, the screen already displayed much of the current state. Anthropic argues that memory was more useful for retaining strategies, plans and rules that the agent had worked out. The company’s database-task discussion similarly emphasizes recording a rule together with evidence for it and warnings about failure cases.
““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, in its definition of agent memory
Limits of the Reported Evaluations
Anthropic’s results come from its own evaluations in a research post, rather than an independently reviewed study. The Pokémon comparison is explicitly based on one run for each model, so the reported badge counts do not show how consistently either system would perform across repeated runs or different starting conditions.
The post’s examples also do not settle how well memory systems work across unrelated tasks, how often saved information becomes stale, or how reliably a review process catches an incorrect note. The source material describes a wiki result of 13% to 48% of graded criteria across 128 sessions, but does not provide enough detail here to assess every evaluation choice or whether the improvement generalizes beyond that task.
Separately, Meyer reports practical problems from his own agents, including crowded indexes, cross-talk between messages and detail lost during handoffs. Those observations are attributed to his experience; they are not Anthropic findings, and the source says Anthropic did not review or sponsor the article.
What Memory Research Must Test
The immediate next step is further testing: repeated runs, more task types and comparisons that make clear which memory components account for changes in performance. Readers should look for independent evaluations and fuller reporting on test setups before treating the examples as general evidence that a particular model or memory architecture is better.
For deployed agents, the practical questions are how notes are selected, when they are retrieved, how they are checked and how users can correct them. The source article recommends date-stamped rules with reasons, but does not report a formal test of that practice. Anthropic’s post describes a proposed memory design and evaluation results; it does not announce a general guarantee that Claude will remember every important instruction across sessions.
Key Questions
What did Anthropic publish?
Anthropic published The components of agent memory on November 6, 2026. The research post describes how Claude agents may save, organize, retrieve and revise information between sessions.
How is agent memory different from a context window?
A context window holds information available during the current session. Agent memory is information saved separately and read in a later session.
Did the Pokémon test prove that fewer memory files are better?
No. Anthropic reported one run each for two models, with different file counts and badge totals. The example suggests that organized, relevant notes may help, but it does not establish a general rule from a single comparison.
What does Anthropic mean by “dreaming”?
It is Anthropic’s name for a between-session process that reviews logs or transcripts and can correct or reorganize saved notes. The term is an analogy for an offline memory-maintenance step, not a claim that models sleep.
Are the results independently verified?
The source describes them as Anthropic’s own evaluations, published in a research post rather than a peer-reviewed paper. Independent confirmation and broader testing remain outstanding.
Source: ThorstenMeyerAI.com