TL;DR
Get everyday essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
Anthropic published a research post on November 6 describing how Claude agents can save and organize information between sessions, then revisit it to correct errors. Its reported tests found better results with a smaller set of focused game notes and a self-built wiki, though the findings include single runs and are not peer-reviewed. The post’s author also describes practical memory problems from operating a small fleet of Claude agents.
Anthropic published a research post on November 6 explaining how AI agents can save information between sessions, organize it for later use and revisit it to correct errors. The post describes memory components for Claude agents and reports evaluations in which a newer model using a smaller set of notes earned more Pokémon Red badges than an older model that created hundreds of files.
Anthropic defines agent memory as knowledge a model gains from interactions or assigned tasks and retains between sessions in material it can read later. It distinguishes that from a model’s context window, which holds information during a session, and from model weights learned during training. The post also treats within-session learning and summaries that replace older conversation turns as separate from persistent agent memory.
The post describes several parts of a memory system: memory notes that record potentially useful information; a salience judgment about what is worth saving; and a process Anthropic calls “dreaming,” in which an agent checks notes against logs and transcripts, corrects mistakes and creates or revises memory. It also discusses structured representations such as linked wiki pages, alongside a hierarchy in which a small set of notes loads automatically and other material is searched only when needed.
In a Pokémon Red evaluation at about 10,000 steps, Anthropic compared one run each of Claude Sonnet 3.7 and Claude Fable 5. Sonnet 3.7 created 413 files across 225 folders and had earned one badge; Fable 5 created nine files in two folders plus a top-level dashboard file and had earned four badges. Anthropic also reports that a self-built wiki raised performance on 25 questions about a company’s documents from 13% to 48% of graded criteria over 128 sessions. These are results reported by Anthropic, not independent findings.
How AI Agents Remember Lessons from Anthropic’s research
Anthropic describes how Claude agents save, organize and revise information between sessions. Its reported tests suggest focused notes and a self-built wiki can help, though the evidence is limited.
What persistent memory does
It carries selected information forward after a session ends.
- Holds information during a session
- Does not automatically carry it into the next session
- Saves selected knowledge for later
- Can be read and revised between sessions
Memory needs upkeep
Anthropic describes a cycle for deciding what to keep and correcting stored information.
Record information that may help with future tasks.
Load a small set automatically and search other material when needed.
“Dreaming” checks notes against logs and transcripts to correct or reorganize them.
The Pokémon Red comparison
At about 10,000 steps, one run per model showed different file counts and badge progress.
| Model | Memory files | Badges earned |
|---|---|---|
| Claude Sonnet 3.7 | 413 across 225 folders | 1 |
| Claude Fable 5 | 9 in 2 folders, plus a dashboard | 4 |
What the evidence can show
The results point to useful design questions for businesses building agents that work across sessions.
Keep memory focused
A smaller set of organized notes may be easier for an agent to use than a sprawling store.
Results are limited
The Pokémon comparison used single runs, and the evaluations were not peer-reviewed or independently verified.
Memory Design Shapes Agent Results
The post focuses on a practical constraint for agents that work over multiple sessions: saved information can guide later decisions, but the agent must decide what to keep, how to find it and when to revise it. A large memory store can become difficult to navigate, while outdated or incorrect notes can carry errors into future work.
Anthropic’s examples suggest that memory quality may matter as much as memory volume in the tasks it tested. The Pokémon comparison associates a compact dashboard and focused notes with more game progress, while the company-document task shows a reported improvement after the agent built a wiki. The results offer design clues, but they do not establish that the same approaches will improve performance across other models or tasks.
Products Worth Considering
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Sky Blue
Samsung Odyssey G7 40-Inch WUHD Curved Gaming Monitor (G75F)
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
As an affiliate, we earn on qualifying purchases.
From Session Context to Saved Notes
A language model’s context window contains material available during a conversation, including user messages and documents. When that session ends, the information does not automatically carry forward unless the system stores it. Persistent agent memory addresses that gap by saving selected information in files or another store for later sessions.
Anthropic’s post applies this idea to Claude and describes memory as a system with both storage and maintenance. Its “dreaming” concept draws on sleep-time compute proposed by researchers at Letta and UC Berkeley, according to the source article. The article’s author says their own small fleet of agents also uses Claude and describes practical problems including full indexes, stale notes, messages crossing between agents and details lost during handoffs. The author says Anthropic did not review or sponsor that account.
““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 Reported Evaluations
The source article says the results come from Anthropic’s own evaluations, published as a research post rather than a peer-reviewed paper, and that several are single runs. The Pokémon comparison involved one run per model, so the reported badge counts do not show how reliably either memory setup performs across repeated attempts.
The post’s reported results do not establish which memory design works best across other models, tasks or real-world deployments. The source material also does not provide enough detail here to assess all evaluation methods or determine how the company-document wiki result would transfer to different collections of documents.
Testing Memory Beyond These Tasks
The next useful step is to see whether the approaches hold up across repeated runs and a wider range of tasks. Anthropic’s post describes components and reports evaluations, but the source material does not identify a schedule for further results or a separate product release tied to the research.
For the author’s own Claude agents, the reported experiments point to ongoing work on keeping indexes manageable, dating notes and recording the reason for rules. Whether those practices reduce errors across a larger set of tasks remains to be established.
Key Questions
What did Anthropic publish?
Anthropic published a research post on November 6, 2026, called “The components of agent memory.” It describes ways Claude agents can save, organize, retrieve and revise information between sessions.
How is agent memory different from a context window?
A context window holds information available to a model during a session. Agent memory is information saved between sessions and read later, according to Anthropic’s definition.
What did the Pokémon Red comparison find?
In one run each at about 10,000 game steps, Claude Sonnet 3.7 created 413 memory files and earned one badge. Claude Fable 5 created nine files in two folders plus a dashboard file and earned four badges. Anthropic reported the comparison; the single runs do not establish typical performance.
What does Anthropic mean by “dreaming”?
It is Anthropic’s term for an agent reviewing logs and transcripts between sessions to correct notes, create new ones or reorganize stored information.
Are the reported findings independently verified?
The source article says the evaluations were published in an Anthropic research post, not a peer-reviewed paper, and that several results came from single runs. It does not describe independent verification.
Source: Thorsten Meyer AI
