🔍 Read the full analysis: What We Learned Running Agents With Memory, Alongside Anthropic’s Research on ThorstenMeyerAI.com
TL;DR
Anthropic published a research post on Nov. 6, 2026, describing how Claude agents can save, organize, retrieve and correct information between sessions. Its post reports results from internal evaluations, including a Pokémon game comparison and a document-question task; a separate account from ThorstenMeyerAI.com describes operational lessons from running Claude agents with memory.
Anthropic published a research post on Nov. 6 explaining how its Claude models can retain and organize information between sessions, as a separate account from ThorstenMeyerAI.com describes lessons from running Claude agents with memory. The research outlines components including saved notes, a process for reviewing them, and layered ways to retrieve information; its reported results come from Anthropic’s own evaluations, not a peer-reviewed paper.
Anthropic defines agent memory as information a model gains through interactions or assigned tasks and retains between sessions in a form it can read later. This is distinct from a model’s context window, which contains information available in the current session. The post also distinguishes memory from changes to model weights, in-context learning during a conversation, and summaries used to make room in a context window.
The research describes several components. Agents can write memory notes about useful information, decide what is salient enough to keep, and later review session records in a process Anthropic calls “dreaming.” That review can correct or rewrite notes. Agents may also organize notes into linked representations such as a wiki, with a small top-level set of information loaded automatically and other material searched when needed.
Anthropic reports that, in a Pokémon Red evaluation at roughly 10,000 game steps, Claude Sonnet 3.7 wrote 413 files across 225 folders and had earned one badge in the cited run. Claude Fable 5 wrote nine files in two folders plus a top-level dashboard file and had earned four badges. The post says these were single runs, so the comparison does not establish that a smaller memory store alone caused the difference.
In another example, Anthropic describes a task involving questions about company documents. A self-built wiki improved the result from 13% to 48% of graded criteria over 128 sessions, according to the post. The source article also reports five practical issues from its own use of Claude agents: indexes can fill, old notes can mislead, messages can cross between tasks, handoffs can lose details, and operating rules are more useful when recorded with a date and rationale. ThorstenMeyerAI.com says Anthropic did not review or sponsor its account.
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 Can Shape Future Agent Work
Persistent memory changes what an agent can do across separate conversations: it can reuse a verified procedure or user preference instead of starting from a blank session. But saved notes also become a source of error if they are stale, poorly organized or treated as proven when they are only guesses. The practical issue is not simply how much an agent can store; it is whether it can select, retrieve and correct information reliably.
Anthropic’s examples offer early evidence about design choices, rather than a broad measure of performance across real-world uses. The Pokémon comparison suggests that recording every event may consume effort without helping, while the document task suggests structured notes can aid repeated question answering. Readers and organizations evaluating agent memory should treat these as reported experimental results, not a guarantee that memory will improve every task.
Popular right now
TCL 75-Inch QM6K QLED Mini LED 4K HDR Smart Google TV, 144Hz
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Eucalyptus Green
De'Longhi All-in-One Espresso and Coffee Maker with 10-Cup Pot and Milk Frother
As an affiliate, we earn on qualifying purchases.
From Chat History to Saved Notes
Traditional chat sessions rely on a context window: a working space holding the current conversation and other material supplied to the model. When the session ends, that information is not necessarily available in a new session. Agent memory adds a separate store that can be carried forward and consulted later.
Anthropic’s post draws an analogy to human memory being consolidated during periods of rest, while describing “dreaming” as an agent’s review of its logs between tasks. The source article links that concept to prior work on “sleep-time compute” by researchers at Letta and UC Berkeley. These are design ideas for systems, not evidence that an AI agent remembers in the same way a person does.
„“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 “The components of agent memory”
Limits of the Reported Evaluations
Anthropic’s findings are presented in a research post, not a peer-reviewed paper, and the source says several evaluations were single runs. The Pokémon results therefore do not show how consistently either model would perform across repeated trials or whether the memory format explains the difference by itself. The reported document-task improvement is tied to a specific question set, scoring method and 128-session evaluation.
The source material does not provide enough detail to compare the approaches on cost, latency, privacy, or performance across a wider range of tasks. It also does not establish how often memory corrections succeed or how systems should handle conflicting notes. ThorstenMeyerAI.com’s operational observations are its own account; Anthropic did not review or sponsor it, according to the publication.
Testing Memory Beyond Examples
The next useful evidence would come from repeated trials and evaluations across more tasks, with clear comparisons of memory formats, retrieval accuracy and the rate of harmful or outdated notes. Anthropic’s post, as described in the source material, does not announce a product launch or a timetable for further results. Readers should watch for additional evaluation details and for evidence about how memory systems handle corrections and task handoffs in routine use.
For teams testing agents now, the reported lessons point to concrete checks: record the source and date of important notes, distinguish tested rules from guesses, and review whether agents retrieve the right information before acting. Those practices do not guarantee accuracy, but they make it easier to identify when a persistent note is helping—or steering a later session wrong.
Key Questions
What did Anthropic publish?
Anthropic published “The components of agent memory” on Nov. 6, 2026. The post describes how Claude agents can save and organize information between sessions, review it later and retrieve it when needed.
How is agent memory different from a context window?
A context window holds information the model can use during a current session. Agent memory is saved for later sessions and can be read back after the current conversation ends.
Did Anthropic prove that fewer memory files lead to better results?
No. The post reports one Pokémon run for each of two models, with different file counts and badge totals. That comparison is suggestive, but it does not establish that a smaller memory store caused the performance difference.
What does “dreaming” mean in the research post?
Anthropic uses the term for an agent reviewing logs and transcripts between sessions. The process can be used to correct notes, create new ones or reorganize information for future use.
What practical risks did the separate account identify?
ThorstenMeyerAI.com said its agents encountered full indexes, stale notes, crossed messages and loss of detail during handoffs. It also said rules worked better when recorded with a date and a reason; these are the publication’s own observations, not Anthropic’s findings.
Source: ThorstenMeyerAI.com
