🔍 Read the full analysis: From Memory Research To Practice: Running AI Agents With Recall 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, 2026, describing how Claude agents can save, organize and revise information between sessions. Its evaluations suggest that selective notes and structured memory can help on some tasks, but the results are company-reported and include single runs.
Anthropic published a research post on Nov. 6 explaining how AI agents can retain and update information between conversations, a capability that can affect whether an agent follows prior instructions or repeats past mistakes. The post, titled “The components of agent memory,” describes memory notes, selection of what to save, later review and organized memory structures; its reported evaluations suggest that more selective, structured notes can outperform a larger collection of files on some tasks.
Anthropic distinguishes an agent’s context window—the information available during its current session—from memory saved for later sessions. Its definition of agent memory is knowledge gained through user interactions or assigned tasks that is retained between sessions in material the model can read again. The post excludes model-weight changes, learning within a single conversation and summaries used to compress an ongoing session.
The proposed components include memory notes, a process for judging which information is likely to matter again, and a later review step Anthropic calls “dreaming.” During that step, an agent can compare saved notes with session logs or transcripts, correct errors and create or revise memory representations. The post also describes a hierarchy in which a short overview is loaded at the start of a session and more detailed material is searched only when needed.
In one Pokémon Red comparison, Anthropic reports that Claude Sonnet 3.7 had created 413 files across 225 folders and earned one badge at about 10,000 steps. Claude Fable 5 had created nine files in two folders and a top-level dashboard file, and had earned four badges. Anthropic also reports that a self-built wiki raised performance on 25 company-document questions from 13% to 48% of graded criteria over 128 sessions. These are results reported by Anthropic, not independent findings.
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 Design Matters
When agents carry information across sessions, the way they store and retrieve it can shape later decisions. A note that preserves a tested preference or task rule may reduce repeated explanations; a stale or incorrect note can instead steer the agent away from a user’s latest instruction. Memory quality is therefore not just storage capacity: selection, organization and correction all matter.
The reported game comparison illustrates a practical trade-off. A large archive can consume effort and make relevant guidance harder to find, while a compact dashboard can surface a current goal and plan. That does not establish that smaller memory is always better: the examples describe particular models, tasks and runs. For people deploying agents, the work points to a need to check what gets saved, how old information is updated and whether important notes can be traced back to evidence.
The source article also describes lessons from its author’s own fleet of Claude-based agents, including crowded indexes, stale notes, crossed messages and details lost during handoffs. Those are the author’s operational observations, not results from Anthropic’s evaluation. They underline why memory systems need maintenance and clear handoff procedures as well as a way to store information.
Popular right now
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
Shark HydroVac Reveal Cordless Wet-Dry Vacuum and Mop, Lilac, WD321
Vitamix 64 oz Low-Profile Blender Container with SELF-DETECT, BPA-Free
As an affiliate, we earn on qualifying purchases.
How Anthropic Defines Agent Memory
The post draws an analogy between agent memory and human memory, suggesting that some sorting and consolidation can happen outside the immediate task. It connects this approach to “sleep-time compute,” an idea previously proposed by researchers at Letta and UC Berkeley. Anthropic presents the analogy as a design framework, not as evidence that AI systems remember in the same way people do.
Its examples cover more than one type of task. In the Pokémon test, the agent had to decide what game information merited a lasting note while continuing to play. In a database task from Continual Learning Bench (CL-Bench) 1.0, questions began in fresh sessions and only a memory folder carried over. Anthropic says some models saved untested guesses, while later models in the comparison more often recorded a rule alongside how it had been checked and warnings about pitfalls.
Anthropic describes these results in a research post, not a peer-reviewed paper. The source article says several results came from single runs. Its author also says the publication did not review or sponsor the article’s account of the author’s own agents.
““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 post does not establish how well the described methods will work across different models, users, tasks or longer deployments. The source material says several results were single runs, and Anthropic’s examples do not by themselves show that the reported gains will generalize. The database result is expressed as the share of graded criteria, not a general measure of agent accuracy.
It is also unclear from the available material how often memories were wrong, how reliably the review step detected errors, or how performance changes as stores grow over time. Anthropic’s post offers design components and reported tests, but the supplied information does not provide an independent replication or a full comparison of costs and failure rates. The practical lessons from the author’s own agents are anecdotal and should not be read as controlled evidence.
Testing Memory Beyond These Examples
Anthropic’s post sets out a framework and evaluations, but the source material does not announce a product release, rollout date or additional test schedule. The next evidence readers would need is repeatable testing across more tasks and runs, including checks of whether agents can identify outdated or unsupported notes and whether memory review improves outcomes without creating new errors.
For teams using agents now, the immediate follow-up is operational: track what the system saves, preserve the basis for important rules, and review notes when instructions or circumstances change. Whether Anthropic will publish further results, or whether independent researchers will reproduce these findings, remains unknown.
Key Questions
What did Anthropic publish?
Anthropic published a research post on Nov. 6, 2026, titled “The components of agent memory.” It describes ways Claude agents can save, organize and revise information between sessions.
How is agent memory different from a context window?
A context window holds information available during a current session. Agent memory is information saved between sessions so the model can read it later.
What does Anthropic mean by “dreaming”?
Anthropic uses “dreaming” for a between-session review in which an agent checks saved notes against logs or transcripts and may correct or reorganize them.
Do the reported results prove that smaller memory stores are better?
No. The Pokémon example favored a smaller, organized store in that comparison, but the reported tests are limited and several were single runs. They do not establish a universal rule for all agents or tasks.
Has Anthropic announced a new memory product?
The supplied source describes a research post and evaluations; it does not report a product launch or rollout schedule.
Source: ThorstenMeyerAI.com
