🔍 Read the full analysis: AI Agent Memory: What Anthropic Studied And What We Experienced 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 Nov. 6, 2026, describing how AI agents can save, organize and correct information between conversations. Its examples suggest that structured, selective memory can help on some tasks, but the results come from Anthropic evaluations, including single runs, and are not peer-reviewed. A separate account from ThorstenMeyerAI.com describes practical memory problems its team encountered with Claude agents.
Anthropic published a research post on Nov. 6, 2026, outlining how AI agents can save information between sessions, decide what to retain and revise notes that may be wrong. The post, titled The components of agent memory, reports results from Anthropic’s own evaluations; it is a research post, not a peer-reviewed paper. ThorstenMeyerAI.com, which says it operates a small fleet of Claude-based agents, separately reports practical problems including stale notes and incomplete handoffs.
Anthropic distinguishes an agent’s context window—the information available during a current session—from memory stored for later sessions. Its definition covers knowledge gained through interactions or assigned tasks and retained as readable context. It excludes changes to model weights, learning within one conversation and summaries used to continue work in a new context window.
The post describes several components: memory notes that record potentially useful information; a salience process for choosing what merits storage; and “dreaming,” in which an agent revisits transcripts between sessions to correct or reorganize notes. It also discusses structured representations such as linked wiki pages and a memory hierarchy, where a short overview is loaded automatically and other information is retrieved when needed.
Anthropic’s Pokémon Red comparison illustrates the trade-off between volume and organization. In one run each, Claude Sonnet 3.7 wrote 413 files across 225 folders and had earned one badge at roughly 10,000 game steps. Claude Fable 5 wrote nine files in two folders plus a top-level dashboard and had earned four badges. The post argues that screen-visible details need not be saved repeatedly, while plans and tested strategies may be useful to retain. These figures describe those particular runs, not a general performance guarantee.
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 Affects Agents
Persistent memory can shape an agent’s decisions after the original conversation is gone. A saved deadline, preference or project rule may help it continue work, but an outdated or mistaken note can also steer later actions in the wrong direction. The issue is not simply how much an agent can store: selection, retrieval and correction all affect what it uses.
Anthropic’s examples point to a practical distinction between raw logs and useful working knowledge. A compact plan or tested rule may be more valuable than a large collection of repeated details. For teams using agents across sessions, the design of memory files and the process for checking them may affect reliability as much as the model’s immediate response.
ThorstenMeyerAI.com says its own Claude agents encountered overfilled indexes, stale information, crossed messages and loss of detail during handoffs. Those are the publisher’s reported experiences, not results of an independent controlled study, and Anthropic did not review or sponsor the account, according to the site.
Products Worth Considering
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Sky Blue
Apple iPad Air 11-inch (M4), 128GB, Wi-Fi 7, Space Gray
DJI Mini 5 Pro Fly More Combo with DJI RC 2
As an affiliate, we earn on qualifying purchases.
From Session Context to Saved Notes
Many chat assistants historically treated each new conversation as a fresh session: the model could use material in its current context but did not automatically retain it for later. Agent memory refers to a separate store that can be read again, such as a folder of files. Anthropic’s post focuses on how that store is written and maintained rather than on changes to the underlying model weights.
The post draws an analogy to human memory processing during rest and cites “sleep-time compute,” a proposal associated with researchers at Letta and UC Berkeley. In Anthropic’s framing, an agent can review session records between tasks, update its notes and prepare information for future use. The article also describes a separate Anthropic task involving questions about company documents: a self-built wiki reportedly raised performance from 13% to 48% of graded criteria over 128 sessions. The source presents this as an evaluation result; it does not establish that the same improvement will occur in other settings.
““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 Evidence
The reported comparisons come from Anthropic’s own evaluations, not a peer-reviewed paper. The Pokémon example is explicitly one run for each model, so it cannot establish typical results across games or tasks. The supplied account also does not provide enough methodological detail to independently assess every comparison or determine how broadly the findings generalize.
It remains unclear how often memory errors occur in real deployments, how well the proposed correction process catches them, and what safeguards prevent an outdated note from being treated as current. The article’s account of its own agents is not a controlled comparison, and the source does not give a measurement protocol for those reported operational problems.
Testing Memory in Daily Use
Anthropic’s post describes approaches and evaluation results but does not announce a release date or a new product in the supplied material. A key next step for assessing the work is whether similar memory structures produce reliable gains across more tasks, models and repeated trials, including cases where saved information becomes outdated.
For users and teams, the immediate practical question is how agents label, retrieve and revise saved notes. ThorstenMeyerAI.com says its own experience favors writing down rules with a date and a reason, but that recommendation is its account rather than a standardized result. Further reporting would need to establish how these practices perform and what controls are available when an agent’s stored memory is wrong.
Key Questions
What did Anthropic publish?
Anthropic published The components of agent memory on Nov. 6, 2026. The research post describes ways agents can store, organize, retrieve and revise information between sessions.
How is agent memory different from a context window?
A context window contains information the model can use during its current session. Agent memory is saved for later sessions and can be read again as context.
Did Anthropic prove that fewer memory files work better?
No. The post reports a Pokémon Red comparison in which one Claude Fable 5 run used fewer files and earned more badges than one Claude Sonnet 3.7 run. Those single runs are illustrative evidence, not proof that fewer files will improve performance generally.
What does Anthropic mean by “dreaming”?
Anthropic uses “dreaming” for an agent’s review of logs and transcripts between sessions, with the aim of correcting mistakes, adding useful notes or reorganizing memory.
What problems did ThorstenMeyerAI.com report?
The site says its Claude-based agents encountered full indexes, stale memory, crossed messages and detail lost during handoffs. These are the publisher’s reported experiences; Anthropic did not review or sponsor the account, according to the site.
Source: ThorstenMeyerAI.com
