🔍 Read the full analysis: A Practical Guide To AI Agent Memory, Informed By Research 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 published a research post on Nov. 6, 2026, describing components of AI agent memory, including saved notes, memory hierarchies and between-session review. Its evaluations report better results in some tasks when memories are selective and structured, but the findings come from Anthropic’s own tests, several of them single runs. A separate article recounts practical problems with agent memory, including stale notes and lost detail during handoffs.
Anthropic published research on Nov. 6 describing how its Claude models store and use information between sessions, offering a practical framework for a growing feature of AI agents: memory that can shape what they do later. The post reports results from Anthropic’s own evaluations, where selective notes and structured memory performed better than large, disorganized collections in some tests; the company’s results are not a peer-reviewed paper, and several comparisons involved single runs.
Anthropic’s post, titled The components of agent memory, distinguishes an agent’s context window from memory saved for later. The context window contains material available during the current session. Agent memory, as the post defines it, is information gained through interactions or assigned tasks that is retained between sessions and saved in a form the model can read again.
The company describes several parts of a memory system: memory notes recording information judged useful; salience, or deciding what is worth keeping; “dreaming,” a between-session process for checking and revising notes against logs or transcripts; and representations such as a linked wiki. A hierarchy can keep a small overview available at the start of each session while leaving detailed notes for targeted search.
In one Pokémon Red comparison, Anthropic reported that Claude Sonnet 3.7 had created 413 files in 225 folders and earned one badge at about 10,000 game steps. Claude Fable 5 had created nine files in two folders, plus a top-level dashboard, and earned four badges. Anthropic presented one run of each model. The post’s interpretation is that recording information already visible in the game can clutter memory, while plans and tested strategies may be more useful to preserve.
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
Memory can affect whether an agent follows a changed deadline, recalls a user’s preference or repeats a mistake across sessions. A system that saves too little may lose useful instructions; one that saves too much can make relevant information harder to find. The article accompanying Anthropic’s post describes practical risks observed in a small fleet of Claude-based agents: stale notes, crowded indexes, crossed messages and incomplete handoffs. These are the author’s reported experiences, not independent evaluation results.
Anthropic’s findings suggest that memory quality is not simply a matter of storage volume. Organizing notes, preserving evidence for a rule and correcting errors could matter as much as what gets saved. For teams using agents on recurring tasks, that makes memory design a reliability issue: a retained error can influence later work, while an incomplete handoff can leave the next session without needed context.
The reported figures should be read as examples from company-run tests, not proof that one memory design will work best across systems. The post’s results may help developers formulate design choices, but they do not establish how memory affects performance in every real-world task.
Popular right now
Vitamix Propel Series 510 Blender, 3 Programs, Slate
Dyson V10 Konical Cordless Stick Vacuum
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
The article explains memory through a comparison: the context window is a desk holding what the model can see now, while long-term memory is a filing cabinet it can consult later. This saved information is distinct from model weights changed during training, knowledge acquired within a single conversation, and summaries used to continue work in a fresh context window.
Anthropic also connects its proposed between-session review process to “sleep-time compute,” an idea previously proposed by researchers at Letta and the University of California, Berkeley. The research post draws an analogy with offline memory processing in people; that analogy is a framing for the approach, not evidence that AI systems remember in the same way humans do.
A second example in the post concerns questions about company documents. Anthropic reported that a self-built wiki improved scores from 13% to 48% of graded criteria across 128 sessions and 25 questions. The measure is the share of graded criteria met, not a percentage increase in general accuracy. The result is from Anthropic’s evaluation and does not, by itself, show that a wiki will produce the same outcome elsewhere.
“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 its research post
Limits of the Published Evaluations
Anthropic published the results as a research post rather than a peer-reviewed paper. The source says several evaluations were single runs, including the Pokémon comparison. It does not establish whether the results would hold across repeated trials, other models, different tasks or independently administered tests.
The post’s examples also leave practical questions open: how often memory should be reviewed, how systems should resolve conflicting notes, and what safeguards prevent sensitive or outdated information from being retained. The accompanying article reports that memory can become stale and that handoffs can lose details, but it provides no quantified failure rates. The source material does not give a full account of the test protocols or enough information to compare performance across all memory approaches.
Testing Memory Beyond These Examples
Anthropic’s post lays out components and test results but does not announce a broader product change or a timetable for further releases. The next evidence readers would need is repeated testing across more tasks, clear comparisons using consistent measures and results from researchers outside the company.
For developers applying these ideas, the practical next step is to check whether saved notes remain accurate and useful over time, rather than judging a memory system by how much it stores. Meyer’s article recommends writing rules with a date and a reason, while Anthropic’s examples point to keeping an index or overview that helps an agent locate relevant details. Those are approaches discussed by the sources, not universal standards.
Key Questions
What does Anthropic mean by AI agent memory?
Anthropic defines it as information Claude gains from interactions or assigned tasks, retains between sessions and saves so it can read the information later.
How is agent memory different from a context window?
The context window holds information available in the current session. Agent memory is saved separately for use in a later session.
Did Anthropic prove that fewer memory files are always better?
No. Its Pokémon example compared one run each of two models, and the post presents a specific evaluation rather than a general rule. The results do not show that fewer files will improve every task.
What is “dreaming” in the research post?
It is Anthropic’s term for a between-session process in which an agent reviews logs or transcripts and may correct, create or reorganize saved notes.
What remains uncertain about agent memory?
The reported tests do not establish how well the methods generalize, how reliably systems correct stale or conflicting notes, or what memory designs work best across different tasks and models.
Source: ThorstenMeyerAI.com
