🔍 Read the full analysis: Exploring AI Agent Memory: What The Research Shows And We Observed 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 research post on Nov. 6 describing how Claude agents can write, organize and revise memories between sessions. Its reported tests suggest selective notes and structured references can help on some tasks, but the results are Anthropic’s own evaluations, including single runs. A separate account describes five lessons from operating a small fleet of Claude-based agents.
Anthropic published a research post on Nov. 6 describing how its Claude models can save and revise information between conversations, bringing attention to a practical problem for AI agents: what they retain can shape how they perform later. The post reports results from Anthropic’s own tests, while a separate account dated Nov. 8 describes operational lessons from a small fleet of Claude-based agents; neither set of observations establishes how memory systems perform across agents generally.
Anthropic defines agent memory as knowledge a model gains from interactions or assigned tasks and retains between sessions in a form it can read later. That differs from a model’s context window, which contains information available during the current session. The post also distinguishes agent memory from changes to a model’s trained weights, in-context learning within one conversation and summaries used to carry work into a new context window.
The post describes several components: agents write memory notes, judge which information is likely to matter again, and can review session records between tasks in a process Anthropic calls “dreaming.” They may also organize notes into structures such as linked wiki pages. A memory hierarchy can load a short overview automatically while leaving detailed notes for targeted searches.
In one Pokémon Red comparison, Anthropic said Claude Sonnet 3.7 created 413 files in 225 folders and had earned one badge at about 10,000 game steps. Claude Fable 5 created nine files in two folders plus a top-level dashboard and had earned four badges. Anthropic also reported that a self-built wiki increased performance on 25 questions about company documents from 13% to 48% of graded criteria over 128 sessions. These figures come from the company’s published evaluations, not an independent study.
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
For users, memory can determine whether an agent carries forward a deadline, preference, tested procedure or project decision. But saved information can also become stale, mistaken or poorly organized. A system that stores too much may make relevant details harder to find; a system that stores too little may repeat errors or lose useful work.
Anthropic’s examples point to a distinction between retaining raw activity and retaining reusable knowledge. The Pokémon comparison suggests that, in that particular run, a compact plan and rules were more useful than a large collection of event files. The wiki result suggests structured references can help with a specific document-question task. Neither example proves that fewer notes or wikis will improve performance across other models, tasks or memory designs.
The practical stakes extend to reliability and oversight. If an agent’s future behavior depends on notes assembled earlier, users and developers need ways to inspect what was saved, identify its source, and correct it. The source account’s operator says its own agents encountered crowded indexes, outdated notes, crossed messages and information loss during handoffs—anecdotal observations that illustrate possible failure modes, not measured rates.
Popular right now
Ninja Luxe Café Mini Espresso and Drip Coffee Machine, Stainless Steel, ES301
TCL 75-Inch QM6K QLED Mini LED 4K HDR Smart Google TV, 144Hz
Hisense 75U6SF Pro 75-Inch U6 Pro Mini-LED 4K UHD Smart Fire TV
As an affiliate, we earn on qualifying purchases.
From Session Context to Saved Notes
Many chat assistants historically began each new conversation without access to earlier exchanges unless users supplied them again or a product provided a separate history feature. Agent memory addresses a different use case: information persists between sessions so a model can resume work without receiving the entire prior conversation.
Anthropic’s post draws an analogy to human memory, where some sorting and consolidation occurs offline. It connects the idea of reviewing records between sessions to “sleep-time compute,” which the source says was proposed by researchers at Letta and UC Berkeley. Anthropic presents the analogy as a design approach for agents, not evidence that machine memory works like human memory.
The accompanying account says its author operates a small fleet of agents running on Claude and that Anthropic did not review or sponsor the account. Its five reported lessons include keeping dated rules with reasons, managing a growing index, checking for stale information, preventing messages from being mixed up and preserving details during handoffs. These are the author’s reported experiences, separate from Anthropic’s evaluations.
““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 Published Results
Anthropic’s results are from its own evaluations, published as a research post rather than a peer-reviewed paper, and the source says several results are single runs. The Pokémon example compares one run each of Sonnet 3.7 and Fable 5, so it cannot isolate memory design from other differences between the models or establish that the result will recur.
The reported wiki improvement applies to 25 questions about company documents and expresses performance as a share of graded criteria across 128 sessions. The source does not provide enough detail here to assess how representative those questions are or how the system would perform on other material. The supplied text also ends partway through its discussion of the database task, leaving that section’s full results and further evaluation details unavailable.
It is also unclear how broadly the author’s five operational lessons apply. The account does not give a sample size, comparison group or quantified error rates for its agent fleet. The research post and the author’s observations should therefore be read as evidence about described tests and experiences, not as a general measure of agent memory reliability.
Further Tests of Agent Memory
The next useful step is further evaluation across repeated runs, different models and varied tasks, with methods that make it possible to separate the effects of model capability from memory organization. Testing should also track whether notes are accurate, whether corrections persist, and how often agents retrieve relevant information rather than outdated or unrelated material.
For practitioners, the account’s lessons suggest recording why a rule was saved and when it was verified, while monitoring indexes and handoffs. Anthropic’s post describes a review-and-rewrite process called dreaming, but the supplied material does not establish whether that approach is available as a standard product feature or how users can apply it. Further implementation details and independent results remain to be seen.
Key Questions
What does Anthropic mean by agent memory?
Anthropic describes it as knowledge retained between sessions in notes or other context that Claude can read later. It is separate from information held only in the current conversation.
What is “dreaming” in the research post?
Anthropic uses the term for an agent reviewing session records between tasks to correct or reorganize its notes, create new memory representations or shape what it loads later.
Did Anthropic prove that smaller memory stores work better?
No. The source reports one Pokémon Red comparison in which Fable 5 used fewer files and had earned more badges than Sonnet 3.7. That result does not establish a general rule across models or tasks.
How strong is the evidence in the post?
The results are Anthropic’s own evaluations, published as a research post rather than a peer-reviewed paper. The source says several were single runs, and the account’s operator separately reports anecdotal experience with a small agent fleet.
What risks come with persistent agent memory?
Saved notes can be out of date, incorrect, hard to find or mixed with other information. The source account reports these as practical problems, while the research post describes reviewing and revising memories as one possible response.
Source: ThorstenMeyerAI.com
