🔍 Read the full analysis: How AI Agent Memory Works: Anthropic’s Research And Our Experience on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Anthropic published a research post on November 6, 2026, describing how Claude agents can save and retrieve information between sessions. Its company-run evaluations suggest that selective, structured memory can help on some tasks, but the results include single-run comparisons and are not peer-reviewed. ThorstenMeyerAI.com also reports operational lessons from its own Claude agents, including risks from outdated notes and incomplete handoffs.
Anthropic published research on November 6 explaining how AI agents can save, organize and revise information between conversations, a capability that can shape what an assistant remembers when a new task begins. The post, titled The components of agent memory, describes memory notes, decisions about what matters, between-session review and layered ways to retrieve saved information; Anthropic’s reported evaluations suggest that useful memory depends on selection and organization, not simply storing more.
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 is different from a model’s context window, which contains information available in the current session, and from model weights set during training. The post also distinguishes memory from summaries used to compress an ongoing conversation.
The proposed system has several parts. An agent writes memory notes about potentially useful information; it judges which events are salient; and it may review session logs later to correct or reorganize what it saved. Anthropic calls that between-session review “dreaming.” The system can also build structured representations, such as linked wiki pages, and use a hierarchy in which a short index is loaded each time while deeper notes are fetched when needed.
Anthropic reported a Pokémon Red comparison using one run each of Claude Sonnet 3.7 and Claude Fable 5, assessed at about 10,000 game steps. Sonnet 3.7 had created 413 files across 225 folders and earned one badge. Fable 5 had created nine files in two folders plus a top-level dashboard and earned four badges. Anthropic argues the smaller, more organized set was more useful because it preserved strategy and plans rather than repeatedly recording information visible on screen. These results are from Anthropic’s own evaluation, not a peer-reviewed 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 Results
Memory gives an agent a way to carry useful information from one session to another, but it also creates a new source of error. An outdated deadline, an untested assumption or a note written for the wrong task can shape later responses. What an agent stores—and what it retrieves—can affect its decisions as much as the information supplied in the current conversation.
Anthropic’s examples make the trade-off concrete: excessive notes can be difficult to navigate, while a small set of tested rules and plans may be easier to use. The post also describes a document-question task in which a self-built wiki raised the share of graded criteria met from 13% to 48% across 128 sessions. That is a result reported by Anthropic for its evaluation; it does not establish that the same approach will improve other systems or tasks.
For users, the practical implication is that persistent memory should not be treated as a perfect record. People may need to check what an agent has saved, correct errors and give it clear instructions about which information should carry forward. For developers, the findings point to retrieval, maintenance and verification as central design problems, not just storage capacity.
Popular right now
Dreame L40 Ultra Gen 2 Robot Vacuum and Mop
De'Longhi All-in-One Espresso and Coffee Maker with 10-Cup Pot and Milk Frother
Vitamix Propel Series 510 Blender, 3 Programs, Slate
As an affiliate, we earn on qualifying purchases.
From Blank Sessions to Saved Notes
Many conversational assistants historically began each new session without access to prior chats, unless a product supplied a separate memory feature. Anthropic’s post addresses a different approach: an agent can write information to files or another store, then read relevant parts in a later session. The source article says the agents in its own operational examples also run on Anthropic’s Claude; it says Anthropic did not review or sponsor that account.
The post draws an analogy with human memory, where some sorting and consolidation appears to happen during rest. Anthropic describes “dreaming” as an agent reviewing transcripts between sessions to fix mistaken notes, derive new information or alter what will be loaded later. The idea builds on “sleep-time compute,” which the post links to work proposed by researchers at Letta and UC Berkeley. The analogy is a design model, not evidence that AI systems remember in the same way people do.
Anthropic also reports a Continual Learning Bench task in which each question began in a fresh session and a memory folder carried over. Its post says some models produced untested guesses, while later models in the comparison more often recorded a rule, evidence for it and warnings about traps. The source article additionally describes practical problems seen in its own agents: an index that fills up, stale notes, messages that cross between tasks and handoffs that lose detail. Those observations are the publisher’s experience, not an independent evaluation.
“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 Evaluations
The reported results do not establish how well these memory designs work across a broad range of models, users or real-world tasks. The Pokémon comparison involved one run per model, and the source presents Anthropic’s findings as a research post rather than a peer-reviewed paper. The cited company-document result is tied to a specific task and 128 sessions; its percentage refers to the share of graded criteria met, not a general measure of agent performance.
It is also unclear from the supplied material how reliably agents identify important information, how often their corrections improve memory, or how the systems handle conflicting instructions and sensitive personal data. Anthropic’s post describes components and evaluations, but the available account does not provide enough detail to resolve those questions. The publisher’s own examples are useful as reported experience, but they are not a controlled comparison.
Testing Memory Beyond Company Tasks
The next useful test is whether these approaches hold up across more tasks and repeated runs, including settings where saved information becomes outdated or conflicts with new instructions. Readers should watch for further evaluations that report task details, comparison methods and failure rates—not only successful examples.
For people using agents with persistent memory, the source article’s practical advice is to treat saved notes as editable working records: check them, date rules and retain reasons or evidence where possible. Anthropic’s post points toward systems that revise memories between sessions, but the available material does not specify a release timetable or confirm that every described component is part of a generally available product.
Key Questions
What does Anthropic mean by AI agent memory?
It means information an agent retains between sessions in a form it can read later, including knowledge gained from interactions or assigned tasks. It is distinct from the current session’s context window and from a model’s training weights.
What is “dreaming” in Anthropic’s account?
Anthropic uses “dreaming” for an agent’s review of logs or transcripts between sessions. The agent may use that review to correct notes, create new ones or change what it retrieves in future sessions.
Did Anthropic prove that fewer memory files are always better?
No. The Pokémon example compared one run each of two models in a particular game setting. Anthropic’s result supports its argument that selective organization can help in that test, but does not show that fewer files will improve every agent or task.
What problems did the source publisher report?
ThorstenMeyerAI.com says its own Claude agents encountered full indexes, stale notes, messages crossing between tasks and detail loss during handoffs. These are the publisher’s observations, not findings from an independent or controlled study.
Source: ThorstenMeyerAI.com
