🔍 Read the full analysis: How Memory Changes AI Agent Runs: What We Observed on ThorstenMeyerAI.com
Get smart everyday buys 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’s November 6 research post describes how Claude agents save, organize and correct information between sessions. Its evaluations suggest that structured, selective memory can help on some tasks, but the results include single runs and do not establish broad performance gains. ThorstenMeyerAI.com also reports five operational problems observed in its own Claude-based agents.
Anthropic published a research post on November 6 describing how its Claude models save, organize and revise information between sessions, offering a look at a feature that can change how AI agents handle repeated tasks. A November 8 report by ThorstenMeyerAI.com pairs the post’s examples with lessons from the publisher’s own Claude-based agents, while noting that Anthropic did not review or sponsor that work.
Anthropic defines agent memory as knowledge a model gains through interactions or assigned tasks and retains between sessions in a form it can read later. That differs from a model’s context window, which holds information available during the current session and may be cleared afterward. The post also distinguishes memory from model training and from summaries used to make room within an ongoing conversation.
The research describes several components: memory notes record information judged useful; salience is the decision about what merits saving; “dreaming” is a between-session process for checking logs and revising notes; and structures such as a wiki can help an agent find related information. Anthropic also describes a memory hierarchy, with a compact overview loaded automatically and other material retrieved only when needed.
In one Pokémon Red comparison, Anthropic reported that Claude Sonnet 3.7 created 413 files across 225 folders and had earned one badge at the reported point in the run. Claude Fable 5 created nine files in two folders plus a top-level dashboard file and had earned four badges. The source says these were single runs, so the comparison is an example, not evidence that the newer model will always perform better.
Anthropic also reported a company-document test in which a self-built wiki increased the share of graded criteria met from 13% to 48% over 128 sessions and 25 questions. On a separate database task, the post describes some models producing untested guesses, while later models included rules, supporting evidence and warnings. These are results from Anthropic’s own evaluations, presented in 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 Agent Memory Can Help—and Fail
Memory can let an agent carry forward a user’s preferences, task progress or lessons from earlier attempts instead of starting every session without that information. The reported examples suggest that how notes are selected and organized may matter as much as how much information is stored: a short plan and tested rules may be more useful than a sprawling record of events.
The trade-off is that saved information can be stale, inaccurate or difficult to retrieve. If an agent carries an old deadline into a new session, it may act confidently on the wrong plan. Anthropic’s account of “dreaming” frames review and correction as part of memory management, while the publisher’s field observations highlight practical risks including crowded indexes, crossed messages and information lost during handoffs.
For readers using AI agents on ongoing work, the findings are a reason to treat memory as an operational record that needs review—not as a guarantee that the system remembers correctly. The reported evaluations do not establish how often these benefits or failures occur in everyday use.
Products Worth Considering
Apple Mac mini Desktop Computer with M2 Pro, 16GB Memory, 512GB SSD, Gigabit Ethernet
Apple Mac mini (2026), M6 12-core CPU and GPU, 24GB Memory, 1TB SSD
Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and 12MP Center Stage Camera (2026, Silver)
As an affiliate, we earn on qualifying purchases.
From Session Context to Saved Notes
Earlier chat assistants generally relied on the information available in the current conversation. A context window functions like a temporary workspace: it can contain the user’s messages, documents and earlier replies, but that material does not necessarily carry into a later session. Agent memory adds a separate store that can be read again when a new task begins.
Anthropic’s post uses an analogy to human sleep and rest to explain why some memory processing might happen between sessions. It also refers to “sleep-time compute,” an idea previously proposed by researchers at Letta and the University of California, Berkeley. The post’s components describe a design approach; they do not mean that an AI model sleeps or remembers in the same way a person does.
ThorstenMeyerAI.com says it operates a small fleet of agents using Anthropic’s Claude and drew five lessons from its experience: indexes can fill, old notes can become dangerous, messages can cross, handoffs can lose detail, and rules are more useful when written down with a date and reason. The publisher says Anthropic did not review or sponsor its observations, which are separate from Anthropic’s published tests.
“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, as quoted in its research post
Limits of the Published Evidence
The reported results have important limits. The Pokémon comparison involved one run each, and the source characterizes Anthropic’s results as its own evaluations, with several examples based on single runs. The company-document figures describe a particular test setup; the material does not show whether the same results would hold across other organizations, tasks or memory designs.
The source does not provide enough detail to assess the full evaluation methods, variation across repeated trials or the rate at which memory errors occur in routine use. It also does not establish how reliably an agent’s between-session review catches incorrect or outdated notes. ThorstenMeyerAI.com’s five lessons are firsthand observations, not a controlled study, and the article says Anthropic did not review or sponsor them.
Testing Memory in Everyday Work
The next useful evidence would come from repeated comparisons across tasks, with clear reporting on how memory is written, retrieved and corrected—and how often errors persist. Readers should also watch for details on whether the approaches described in Anthropic’s post appear in specific Claude products or agent workflows; the supplied material does not identify a product rollout or timetable.
For now, the published post and the publisher’s account offer design lessons rather than a general performance guarantee. Teams adopting agents can track whether saved notes remain accurate, whether handoffs preserve key details and whether the agent can show the evidence behind a stored rule. No further release or milestone is specified in the source material.
Key Questions
What does AI agent memory mean here?
Anthropic describes it as information an agent retains between sessions in notes or other saved context that it can read later. It is distinct from the information visible only in the current conversation.
What did the Pokémon test report?
Anthropic reported one run each: 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. These single runs do not establish typical results for either model.
What does Anthropic mean by “dreaming”?
It is the post’s term for an agent reviewing logs and transcripts between sessions to check, correct or reorganize saved memory. The term is an analogy for a processing step, not a claim that the model sleeps like a human.
Did Anthropic verify the publisher’s five lessons?
No. ThorstenMeyerAI.com says its agents use Claude, but also says Anthropic did not review or sponsor the publisher’s observations. They are separate from the company’s evaluations.
Do the results prove that memory improves every agent?
No. The examples indicate that memory structure can matter on particular tests, but the source reports limited evaluations, including single runs. Broader performance and error rates remain unclear.
Source: ThorstenMeyerAI.com
