🔍 Read the full analysis: The Ideas Behind AI Agent Memory—and What Happened When We Tried Them 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 November 6, 2026, outlining how Claude agents can save, organize and revise information between sessions. The post reports better results in some tests with compact, structured memories, but the evaluations were conducted by Anthropic and include single-run comparisons. A separate account of the publisher’s own agents describes practical problems including outdated notes and lost handoff details.
Anthropic published a research post on November 6 describing how its Claude models can save, organize and correct information between sessions, an effort aimed at reducing the need for users to repeat details. The post reports that structured memory performed better than a large collection of scattered files in some tests, but its results come from Anthropic’s own evaluations, including several single runs, rather than an independent or peer-reviewed study.
Anthropic’s post, titled “The components of agent memory,” distinguishes working memory—the information available in a model’s current context window—from memory saved for later sessions. The company defines agent memory as knowledge gained from interactions or assigned tasks that the model retains between sessions in a form it can read again. This is separate from changing model weights, learning within a single conversation or summarizing older turns to make room in a context window.
The system Anthropic describes has several parts: an agent decides what is worth saving, records notes, and can later review those notes against transcripts to correct or reorganize them. Anthropic calls this review step “dreaming.” It also describes memory representations such as a linked wiki, and a hierarchy in which a small overview is loaded automatically while more detailed information is retrieved when needed.
In one Pokémon Red comparison, Anthropic examined single runs of Claude Sonnet 3.7 and Claude Fable 5 at about 10,000 game steps. 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 said the smaller collection focused on goals, plans and rules, while information visible on the screen each turn did not need to be saved.
Another reported test used questions about a database, with each question starting in a fresh session and a memory folder carrying over. Anthropic said a self-built wiki raised performance on 25 questions about company documents from 13% to 48% of graded criteria over 128 sessions. The post also describes differences among model versions: some saved untested guesses, while later versions more often recorded rules alongside evidence and warnings about traps.
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 Structured Memory Matters
Persistent memory can change an assistant from a tool that responds only to the current conversation into one that carries selected information forward. That may reduce repeated instructions and help agents handle work that spans sessions. The tests suggest that what an agent chooses to remember, and how it organizes those notes, can matter as much as how much information it stores.
But saved notes also create a new source of error. An outdated deadline, an unsupported guess or information assigned to the wrong person can influence later actions. The research therefore matters not just as a performance question, but as a practical question of how memory is checked, retrieved and corrected. Anthropic’s examples point to possible methods; they do not establish how reliably those methods work across products or real-world tasks.
A separate account accompanying the research describes five problems encountered by the publisher’s own agents: an index that fills up, stale memories, messages crossing, handoffs losing detail and rules that work better when recorded with a date and reason. The account says those agents use Claude and that Anthropic did not review or sponsor it. These are the publisher’s reported experiences, not findings from Anthropic’s evaluations.
Products Worth Considering
Apple MacBook Neo 13-inch Laptop with A18 Pro Chip, 8GB Memory, 256GB SSD, Liquid Retina Display, and 1080p Camera – Blush
Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Midnight
Apple MacBook Air 13.6-inch with M5 chip, 16GB Memory, 512GB SSD, Midnight
As an affiliate, we earn on qualifying purchases.
From Chat Context to Saved Notes
For a chat without persistent memory, the model can use the messages and materials in its current context window, but that working context does not automatically carry into a new session. In Anthropic’s description, an agent-memory system saves selected information in files or other structures that it can read later. The distinction matters: a model may have a large context window and still lack a durable record of what a user said in an earlier conversation.
Anthropic’s approach draws on the idea that memory can be reviewed outside the immediate task, which the post relates to human memory consolidation during rest. The post also builds on the concept of sleep-time compute, previously proposed by Letta and UC Berkeley researchers. Anthropic presents notes, review, linked representations and a layered retrieval system as components that can work together, rather than treating memory as a complete transcript loaded into every interaction.
“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 November 6 research post
Limits of the Published Tests
The reported results do not establish that compact memory will outperform larger stores in every task. Anthropic says several evaluations were single runs, and the comparisons involve particular models, tasks and settings. The Pokémon result, for example, compares one run of each model; it cannot by itself show that memory design caused the difference in badges earned.
The database figures are reported by Anthropic, and the post is a research post rather than a peer-reviewed paper. The available account does not establish how well the methods generalize to other models, longer deployments or memory containing conflicting and sensitive information. It also does not provide a broad independent assessment of how often the “dreaming” process detects and corrects false notes.
The publisher’s account of its own agents offers practical observations but does not give enough detail to treat them as controlled test results. It is also unclear from the supplied material what safeguards or user controls are available across Claude products that use persistent memory, or how users would inspect, edit or delete all saved information.
Further Tests and User Controls
The next evidence to watch for is whether Anthropic or independent researchers publish repeated evaluations across a wider range of tasks, with enough detail to compare memory structures and correction methods. Replicated tests could help distinguish a consistent benefit from results tied to a particular model or run.
For users and organizations considering persistent-memory agents, the practical questions remain how notes are reviewed, how stale or mistaken information is removed, and what controls people have over stored data. Anthropic’s post describes possible system components, but the material provided does not set out a product-wide schedule for new features or further research.
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 models 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 the model during a session. Agent memory is selected information saved so the model can read it in a later session.
What does Anthropic mean by “dreaming”?
Anthropic uses “dreaming” for a between-session review in which an agent checks notes against logs or transcripts and can correct or reorganize its stored memory.
Do the reported results prove that smaller memory stores are better?
No. Anthropic reports favorable results in specific evaluations, including a single-run Pokémon comparison. Those results do not show that compact memory will perform better across all models and tasks.
What problems can persistent agent memory create?
Notes can become outdated, contain unsupported guesses or be associated with the wrong message or task. The publisher’s account also reports lost detail during agent handoffs; those are its own observations, not controlled findings from Anthropic’s tests.
Source: ThorstenMeyerAI.com
