🔍 Read the full analysis: How We Ran AI Agents With Memory—and What We Learned on ThorstenMeyerAI.com
Get hardware and tech essentials 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 Nov. 6, 2026, describing how Claude agents can save, organize and revise information between sessions. Its examples suggest that smaller, structured memories can outperform sprawling notes in some tests, while a separate account of running Claude agents points to risks including stale information and lost handoff details.
Anthropic published a research post on Nov. 6, 2026, describing how Claude agents can save information between sessions, organize it for later use and revise notes against earlier records. The post reports results from several company-run evaluations; a separate account by Thorsten Meyer says a small fleet of Claude agents also exposed practical problems, including stale notes, crowded indexes and incomplete handoffs.
Anthropic defines agent memory as knowledge gained through interactions or assigned tasks that a model retains between sessions in a form it can read later. That is different from a model’s context window, which holds material available during a current session, as well as from changes to model weights or summaries that replace earlier conversation turns. The research post, “The components of agent memory,” describes notes, decisions about what is worth saving, later review of records, and structures such as linked wiki pages.
In one Pokémon Red evaluation, Anthropic compared one run each of Claude Sonnet 3.7 and Claude Fable 5 at about 10,000 game steps. The post says Sonnet 3.7 created 413 files across 225 folders and had earned one badge. Fable 5 created nine files in two folders plus a top-level dashboard file, and had earned four badges. These are results from single runs, not a broad benchmark establishing that one memory design is generally superior.
Anthropic also describes a database task in Continual Learning Bench 1.0, where each question began in a fresh session and only a memory folder carried forward. The company says some models produced untested guesses, while later models recorded rules along with evidence and warnings about traps. In a separate company-document task, the post says an agent-built wiki raised scores from 13% to 48% of graded criteria across 25 questions and 128 sessions. These findings are Anthropic’s own evaluations, published as 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
Why Structured Agent Memory Matters
Memory can let an agent carry useful information from one session to the next, reducing the need for a user to repeat details and giving the system a way to build on earlier work. The examples suggest that simply saving more is not necessarily better: a compact summary with a current goal and tested rules may be easier to use than a large collection of disconnected files.
The potential benefit comes with a reliability problem. If an agent saves an incorrect assumption or fails to update an old instruction, it may apply that information later as though it were still accurate. Meyer’s account of his own agents flags stale memory, index growth, messages crossing and detail lost during handoffs. Those observations are his operational experience, not findings independently verified by Anthropic.
For readers using agents on ongoing work, the distinction matters because saved notes can shape future outputs without being visible in every session. Memory design is therefore not just a storage question: it affects what the agent retrieves, what it treats as established and whether a user can identify the source or age of a remembered detail.
Products Worth Considering
Apple MacBook Pro 14.2-inch with M5 Max, 18-Core CPU, 20-Core GPU, 48GB Memory, 2TB SSD, Space Black
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Sky Blue
Apple MacBook Pro 16.2-inch with M5 Pro, 18-core CPU, 20-core GPU, 64GB Memory, 1TB SSD, Space Black
As an affiliate, we earn on qualifying purchases.
How Agent Memory Differs From Chat
A conventional chat session gives a model access to the messages and other material in its current context window. When that session ends, the model does not automatically carry the conversation forward unless a system separately saves information. Anthropic’s post focuses on that separate layer: information written to storage and read in a later session.
The post uses a comparison with human memory, describing a division between recording experiences and later sorting or consolidating them. It calls the between-session review process “dreaming,” a term for checking notes against logs or transcripts and then correcting or reorganizing them. Anthropic links the idea to “sleep-time compute,” previously proposed by researchers at Letta and UC Berkeley.
The source article also reports five lessons from Meyer’s own agents: indexes can fill, old notes can become dangerous, messages can cross, handoffs can lose detail, and rules are more useful when recorded with a date and reason. Meyer says the agents run on Claude and that Anthropic did not review or sponsor his account. The two bodies of material should be kept distinct: Anthropic’s published evaluations are company-reported tests, while the five lessons are an operator’s account.
“The index fills up, stale memory is dangerous, messages cross, handoffs lose detail, and rules work best as memory written down with a date and a reason.”
— Thorsten Meyer, describing the results of his own agents
Limits of the Reported Tests
The evidence described in Anthropic’s post has important limits. The Pokémon comparison involved one run per model, and the source says several of the reported results were single runs. The article does not provide enough detail to establish whether the same differences would hold across repeated tests, other tasks or other agent systems.
The database-task account names model versions and describes differences in their notes, but the supplied material does not give numerical scores for that comparison. The wiki result is reported as a change in the share of graded criteria met, not as a general measure of agent accuracy. It is also unclear from the supplied account how the evaluation was repeated, how grading was conducted or how the results compare with alternative memory designs.
Anthropic’s post is described as a research post, not a peer-reviewed paper. The source material does not say whether the company plans to release the underlying data or full evaluation methods. Meyer’s operational observations are not a controlled study, and the article does not quantify how often the reported memory problems occurred.
What Further Testing Could Show
The immediate next step for readers is to distinguish the reported examples from evidence of broad performance. Repeated evaluations across tasks and runs could help show whether compact files, linked wiki structures and review between sessions reliably improve results. The supplied source does not identify a date for further Anthropic research, a product launch or a follow-up benchmark.
For agent operators, the account points to practical checks rather than a guaranteed fix: track when a note was written, preserve why it was saved, and compare important remembered claims with the original record before relying on them. Those are lessons Meyer reports from his own use, not a formal recommendation issued by Anthropic.
How agents should decide what to retain, how they should signal uncertainty in saved notes and how users can inspect or correct those notes remain open questions in the material provided. Until more comparable results are available, the reported gains should be read as examples of what Anthropic’s systems did in particular evaluations, not as proof that memory improves every agent or task.
Key Questions
What did Anthropic publish?
Anthropic published a research post on Nov. 6, 2026, called “The components of agent memory.” It describes how Claude agents can save information, choose what to retain, organize notes and review them between sessions.
How is agent memory different from a context window?
A context window contains information a model can use during its current session. Agent memory is information saved to storage and made available in a later session, according to Anthropic’s definition quoted in the source article.
Did Anthropic prove that fewer memory files are always better?
No. The Pokémon example compared one run each of two models. It suggests that a smaller, organized memory worked better in that test, but it does not establish a general rule across other runs, models or tasks.
What problems did Meyer report with his agents?
Meyer says his agents’ memory indexes can grow, old information can become stale, messages can cross and handoffs can lose detail. He also says rules were more useful when saved with a date and reason. These are his reported observations, not results from Anthropic’s evaluation.
What does Anthropic mean by “dreaming”?
Anthropic uses “dreaming” for a between-session process in which an agent reviews logs or transcripts and may correct notes, add information or reorganize memory for future use. The supplied material does not establish how consistently this process prevents errors.
Source: ThorstenMeyerAI.com
