🔍 Read the full analysis: AI Memory In Practice: What We Learned From Running Agents 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 November 6 research post describing how its Claude models write, organize and revise information saved between sessions. Its company-run tests suggest that structured, selective memory can help on some tasks, but the results include single runs and do not establish how the methods perform broadly.
Anthropic published a research post on November 6 outlining how its Claude models save information between conversations, organize it for later use and revise it when it may be wrong. The post describes agent memory as a system distinct from a model’s immediate context window, and reports company-run tests in which smaller, structured collections of notes performed better than a large, unorganized store on specific tasks.
Anthropic’s post, titled The components of agent memory, describes several parts of the process: memory notes that record potentially useful information; a salience judgment about what merits saving; an offline review step Anthropic calls “dreaming”; and representations such as linked wiki pages that make stored information easier to navigate. The system can also use a hierarchy, with a small set of notes loaded at the start of a session and other material retrieved only when needed.
The company reports a comparison from a Pokémon Red task. At roughly 10,000 steps, one run with Claude Sonnet 3.7 had produced 413 files across 225 folders and earned one badge. One run with Claude Fable 5 had produced nine files in two folders, plus a top-level dashboard file, and earned four badges. Anthropic says the newer model’s notes emphasized plans and rules rather than repeatedly recording information visible on screen. These are results from individual runs, not a broad statistical comparison.
Another example involved questions about company documents across 128 sessions. Anthropic says a self-built wiki increased performance from 13% to 48% of graded criteria on 25 questions. The source account also describes practical issues from running a small fleet of Claude agents: indexes can fill, saved notes can become stale, messages can cross between agents and handoffs can lose detail. Its author says dated rules with an explanation of why they were added worked better in that setup.
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 Agent Memory Quality Matters
Memory affects what an agent carries into a later task, so a saved mistake can shape future answers just as a useful instruction can. Persistent notes can reduce repeated explanations, but the examples also show why simply storing more information is not a reliable solution: an agent must select relevant details, retrieve them appropriately and correct outdated or inaccurate entries.
For users and organizations deploying agents across sessions, the distinction matters operationally. A clear plan or tested rule may be more useful than a large archive of conversation fragments. At the same time, Anthropic’s examples are company-reported evaluations, and the practical observations in the source article come from one operator’s own agents. They offer evidence about possible approaches, not a guarantee that a particular memory design will improve results in other systems.
Products Worth Considering
Apple MacBook Air 13.6-inch with M5 chip, 16GB Memory, 512GB SSD, Midnight
Apple MacBook Neo 13-inch Laptop with A18 Pro Chip, 8GB Memory, 256GB SSD, Liquid Retina Display, and 1080p Camera – Blush
FLEXISPOT Adjustable Workbench with Hand Crank, 47.2 x 23.6 Rubberwood Top
As an affiliate, we earn on qualifying purchases.
From Chat History to Saved Notes
A model’s context window is the information available to it during its current session: messages, documents and earlier replies. The source article compares it to a desk that is cleared when the session ends. Agent memory, by contrast, is information saved between sessions and made available later. Anthropic’s definition covers knowledge gained through interactions and assigned tasks that is retained as readable context.
The post distinguishes this from changes to model weights during training, learning within one conversation, and summaries that replace earlier turns to make room in a context window. It also draws on the idea that useful information can be processed between active sessions. Anthropic connects its “dreaming” step to sleep-time compute, an approach proposed by researchers at Letta and UC Berkeley. Its post is a research account of the company’s methods and evaluations, not a peer-reviewed paper.
““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 available source describes Anthropic’s findings as company-run evaluations, with several results coming from single runs. It does not establish whether the Pokémon comparison would hold across repeated trials, other games or different agent tasks. The reported improvement on document questions also applies to the stated set of 25 questions and 128 sessions; the source does not provide enough detail here to assess how broadly that result generalizes.
It remains unclear how often the memory-review process catches false or outdated notes in routine use, how much time or computing it adds, and how the described methods compare with other memory systems under the same conditions. The operational problems in the source article are the author’s reported experience, not independent measurements. Anthropic did not review or sponsor that account, according to its author.
What Further Evaluations Could Show
The immediate next step for readers evaluating these claims is to distinguish the post’s reported examples from evidence across repeated, comparable tests. Anthropic’s account outlines a design—selective notes, structured representations and later correction—but the examples provided do not settle how well that design works across settings.
Teams using persistent agent memory can monitor whether notes remain accurate, whether agents retrieve the right information and whether handoffs preserve instructions. The source material gives no date for a follow-up evaluation or further Anthropic announcement. Broader performance claims will depend on additional testing, including how memory systems handle stale information and errors over time.
Key Questions
What did Anthropic announce?
Anthropic published a research post on November 6, 2026, describing how Claude agents can save, organize, retrieve and revise information between sessions. It is a research post, not an announcement of a new consumer feature in the source material.
How is agent memory different from a context window?
A context window holds information the model can see during its current session. Agent memory is saved between sessions and can be read in a later one.
What does Anthropic mean by “dreaming”?
Anthropic uses “dreaming” for an offline review of logs and transcripts, where an agent can check its notes, correct mistakes and create or revise memory representations.
Do the reported results prove that less memory is always better?
No. The Pokémon example compares one run per model in the reported test. It suggests that structured, selective notes helped in that example, but does not establish a general rule across tasks or repeated trials.
What practical problems did the source author report?
The author described indexes filling up, stale notes, messages crossing between agents and detail being lost during handoffs. Those are the author’s observations from a small fleet of Claude agents, not independently verified findings.
Source: ThorstenMeyerAI.com
