🔍 Read the full analysis: AI Agent Memory, Explained Through Research And Hands-On Experience on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Anthropic published a research post on November 6, 2026, describing how AI agents can save information between sessions, organize it for later use and revise it against past interactions. The company reported better results in some tests when agents kept concise, structured notes, but its findings are company-run evaluations, including several single runs, not peer-reviewed results.
Anthropic published a research post on November 6 explaining how its Claude models can save and revise information between sessions, a capability that may help agents carry useful knowledge across tasks instead of starting each conversation with a blank slate. The post describes memory notes, selection of what to retain, later review and structured files, and reports results from Anthropic-run evaluations; it is not a peer-reviewed paper.
Anthropic distinguishes an agent’s working context—the messages and information available during a session—from memory that is saved and read in later sessions. Its definition covers knowledge gained through user interactions and assigned tasks that the agent retains as readable context. It excludes changes to model weights, learning within a single conversation and summaries that replace earlier messages to extend a session.
The research post describes several connected components: memory notes record information an agent judges worth keeping; salience is the process of choosing what may matter again; and “dreaming” is Anthropic’s name for a between-session review of logs and transcripts to correct or reorganize notes. Agents can also create memory representations, such as linked wiki pages, and use a hierarchy in which a short overview is loaded automatically while other files are retrieved as needed.
In a Pokémon Red test, Anthropic compared one run each of Claude Sonnet 3.7 and Claude Fable 5 at about 10,000 game steps. The company said Sonnet 3.7 created 413 files across 225 folders and had earned one badge; Fable 5 created nine files in two folders and a top-level dashboard, and had earned four. Anthropic also reported that a self-built wiki raised performance on 25 questions about company documents from 13% to 48% of graded criteria over 128 sessions. These are results reported by Anthropic, not independent findings.
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
Why Memory Design Changes Agent Work
Memory affects whether an agent can reuse a tested rule, honor a user’s updated instruction or continue a multi-session task without repeating earlier work. Saving more is not automatically better: the Pokémon comparison suggests that a large collection of fragmented files can be less useful than a smaller, organized record that highlights goals and strategies. The comparison is limited to one run per model, so it does not establish a general performance rule.
Memory also creates a reliability problem. A stored note can be mistaken, outdated or stripped of the context that made it accurate. Anthropic’s “dreaming” approach addresses this by checking notes against earlier records, while the post’s database-task examples emphasize documenting how a rule was tested and what exceptions apply. For people using agents, this means memory quality depends on correction and retrieval, not just persistence.
The source article also describes hands-on experience operating a small fleet of Claude-based agents. Its author reports recurring problems with an overfilled index, stale notes, messages crossing between agents and detail lost during handoffs. The account is practical experience rather than a controlled study; it points to risks users may want to monitor, but does not quantify how often they occur.
Products Worth Considering
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Sky Blue
Jackery Explorer 300 Portable Power Station with 40W Solar Panel
Jackery Explorer 1000 v2 Portable Power Station, 1070Wh, 1500W
As an affiliate, we earn on qualifying purchases.
From Blank Sessions to Saved Notes
For an assistant without persistent memory, each new conversation begins without the previous session’s content. A context window functions as the information available at that moment; once a session ends, details do not automatically carry forward. Agent memory adds a separate store that can be read in a later session, which is why its contents can influence future answers and actions.
Anthropic’s post draws an analogy to human memory consolidation, describing some sorting as happening offline during sleep or rest. It connects its proposed between-session review to “sleep-time compute,” an idea previously proposed by researchers at Letta and the University of California, Berkeley. The analogy describes a design approach, not evidence that an AI agent remembers in the same way a person does.
The article supplying this account says its author runs Claude-based agents and that Anthropic did not review or sponsor the article. It also reports a separate document-question test: across 128 sessions and 25 questions, a self-built wiki increased the share of graded criteria met from 13% to 48%. The available material does not provide enough detail to treat that result as independently verified or broadly representative.
“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 research post
Limits of the Reported Tests
The results come from Anthropic’s own evaluations, not a peer-reviewed paper or independent replication. The Pokémon comparison used one run per model, and the reported figures do not show how consistently the outcomes would recur across different tasks, prompts or memory setups. The source material also does not provide full evaluation methods or uncertainty estimates for every result.
It remains unclear how well these memory methods generalize beyond the tested tasks, how often automated review catches an incorrect note, or what safeguards prevent one user’s or agent’s information from being applied in the wrong setting. The hands-on reports of stale memories and crossed messages are attributed observations, not measured rates. The relative costs of storing, reviewing and retrieving notes are also not specified.
What to Watch in Future Evaluations
The immediate next step is further testing: the reported comparisons need repeated runs and evaluation beyond Anthropic’s own examples to establish how reliable the gains are. Readers should look for details on methods, reproducibility and performance across tasks, as well as evidence that memory corrections reduce errors without removing useful information.
For teams using agents now, the practical questions are whether saved notes have clear dates and sources, whether agents can retrieve only relevant records, and how handoffs are checked for lost or misdirected details. Anthropic’s post describes a set of design components; the material provided does not announce a new product release or give a timetable for further research.
Key Questions
What does AI agent memory mean in Anthropic’s post?
It means information an agent saves between sessions as readable context, including knowledge gained from interactions and assigned tasks. It is distinct from the information available only in the current conversation.
How is persistent memory different from a context window?
A context window holds what the model can use during a particular session. Persistent memory is stored separately and may be read in a later session, allowing selected information to carry forward.
What did Anthropic report in its Pokémon test?
In one run per model, Anthropic said Sonnet 3.7 created 413 files across 225 folders and earned one badge, while Fable 5 created nine files in two folders plus a dashboard and earned four badges. The result is a limited company-reported comparison, not proof that the newer approach will perform better in every task.
What does Anthropic mean by “dreaming”?
It is the post’s term for reviewing logs or transcripts between sessions to correct, add or reorganize memory before the agent uses it again.
Are the reported results independently verified?
The source describes them as Anthropic’s own evaluations. It does not identify them as peer-reviewed or independently replicated findings, and several comparisons involved single runs.
Source: ThorstenMeyerAI.com
