🔍 Read the full analysis: What Does It Take For AI Agents To Remember? Research And Lessons on ThorstenMeyerAI.com
Get hardware and tech essentials 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 describing components of AI agent memory, including memory notes, salience, offline review and layered representations. Its reported tests suggest that compact, structured notes can outperform sprawling files in specific tasks, but the results come from Anthropic evaluations, several based on single runs.
Anthropic published a research post on November 6 describing how its Claude models can save, organize and revise information between conversations, a capability that affects whether AI agents can use lessons from earlier tasks. The post, titled The components of agent memory, reports results from company evaluations and presents a framework for separating short-term context from persistent notes.
Anthropic defines agent memory as knowledge gained through interactions or assigned tasks that a model retains between sessions in a form it can read later. That differs from a model’s context window, which contains information available during the current conversation and does not, by itself, carry details into a later session. The company also distinguishes memory from changes to model weights, learning within one conversation and summaries used to continue work in a new context window.
The post describes several components. An agent writes memory notes about information it judges likely to matter again; this judgment is called salience. Anthropic uses “dreaming” for a separate review process in which an agent checks transcripts and logs between sessions, corrects or replaces notes, and creates new representations. Those representations might include a wiki, while a memory hierarchy can keep a small overview available by default and retrieve more detailed material only when needed.
In a Pokémon Red test, Anthropic compared one run each of Claude Sonnet 3.7 and Claude Fable 5 at about 10,000 game steps. Sonnet 3.7 created 413 memory files across 225 folders and had earned one badge. Fable 5 made nine files in two folders, plus a top-level dashboard, and had earned four badges. Anthropic says the newer model’s notes focused more on strategy and plans, rather than repeatedly recording information visible on screen. These figures describe the reported runs, not a broad benchmark result.
In another test, a self-built wiki helped an agent answer 25 questions about company documents. Across 128 sessions, the post reports that the wiki raised scores from 13% to 48% of graded criteria. Anthropic also describes database tasks in which questions began in fresh sessions and only the memory folder carried over. It says some models made untested guesses, while later models 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 Memory Design Changes Agent Performance
Persistent memory can let an agent carry forward user preferences, task-specific rules and lessons from earlier attempts instead of starting each session without that information. That could make systems more useful for work spanning multiple conversations, such as maintaining a project or answering questions across a large document collection.
The results also point to a trade-off: retaining more information is not automatically better. A crowded store can make useful knowledge harder to find, while a short, structured note may help an agent act more effectively. But memory introduces another risk: a mistaken note can shape later answers. Anthropic’s description of review and correction addresses that problem in principle; the published examples do not establish how reliably agents will detect errors across everyday use.
The distinction matters to users because an agent’s remembered notes can influence future responses without being visible in the current conversation. Readers should treat the reported performance as evidence from particular company tests, not as proof that every agent with persistent memory will be accurate or dependable.
Products Worth Considering
Apple MacBook Pro 16.2-inch with M5 Pro, 18-core CPU, 20-core GPU, 64GB Memory, 1TB SSD, Space Black
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Sky Blue
Apple MacBook Pro 14.2-inch with M5 Max, 18-Core CPU, 20-Core GPU, 48GB Memory, 2TB SSD, Space Black
As an affiliate, we earn on qualifying purchases.
From Blank Sessions to Saved Notes
For much of the history of chat assistants, each new conversation began without access to the previous session unless information was supplied again or a separate memory feature was used. A context window is the material the model can work with at that moment; persistent agent memory is information saved for possible use later.
Anthropic’s post draws an analogy to human memory, where some sorting and consolidation is associated with rest, and connects its “dreaming” concept to sleep-time compute proposed by researchers at Letta and the University of California, Berkeley. The analogy describes a design idea, not evidence that an AI system remembers in the same way people do.
The article’s source also reports practical lessons from the publisher’s own small fleet of agents running on Claude: indexes can fill up, stale notes can mislead, messages can cross, and handoffs can lose details. It says the publisher was not sponsored by or involved in Anthropic’s research. These observations are operational experience, separate from Anthropic’s evaluations.
“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 Published Evaluations
The results are from Anthropic’s own evaluations, published as a research post rather than a peer-reviewed paper. The post says several results are single runs; the Pokémon comparison, for example, includes one run per model. The material does not establish whether the differences would hold across repeated trials, other tasks or agents made by other developers.
It is also unclear how often memory-review systems catch false or outdated notes in routine deployment, how users will be shown what an agent has stored, and what controls they will have to inspect or remove it. The source gives practical warnings about stale information and confused handoffs but does not provide rates for those failures or a complete account of privacy safeguards.
Testing Memory Beyond These Tasks
The immediate next step is broader evaluation: repeating these tests, comparing memory approaches across more tasks, and measuring both answer quality and the frequency of persistent errors. Anthropic’s post lays out components and examples, but the supplied material does not announce a release date, product change or follow-up study.
For users and organizations considering memory-enabled agents, the practical questions are whether saved information can be inspected, corrected and deleted; how the system decides what to retain; and whether it can show the evidence behind a remembered rule. Until those details and broader results are available, the reported gains should be understood within the specific evaluations described.
Key Questions
What does AI agent memory mean?
In Anthropic’s definition, it is knowledge retained between sessions in a form the agent can read later. It is separate from the context available only during the current conversation.
What did Anthropic’s Pokémon test find?
In one run per model at about 10,000 steps, Sonnet 3.7 made 413 files across 225 folders and had one badge; Fable 5 made nine files in two folders plus a dashboard and had four badges. Anthropic presents this as an example, not a repeated benchmark result.
What does Anthropic mean by “dreaming”?
It is the proposed between-session review of logs and transcripts, where an agent can check, revise or add memory notes before later sessions.
Do these results prove that AI agents will remember accurately?
No. The post reports company evaluations, several involving single runs. It does not establish how reliably memory systems will avoid or correct errors across other tasks and real-world use.
Source: ThorstenMeyerAI.com
