🔍 Read the full analysis: What We Learned About Memory By Running AI Agents 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 November 6, 2026, describing how Claude agents can save, organize and revise information between sessions. Its reported tests suggest that smaller, structured memory stores can outperform sprawling notes in some tasks, but the results come from Anthropic’s own evaluations and include single-run comparisons.
Anthropic published a research post on November 6, 2026, describing how its Claude models can save, organize and correct information between sessions—a development that could help AI agents carry useful knowledge from one task to the next. The post reports that memory design affected performance in several evaluations, although Anthropic’s results are company-run tests, not peer-reviewed findings.
Anthropic defines agent memory as information a model gains from interactions or assigned tasks and retains between sessions in a form it can read later. That differs from a model’s context window, which contains material available during the current conversation and does not, by itself, preserve information for a future session. The company also distinguishes saved memory from changes to model weights, learning within one conversation, and summaries used to make room in a context window.
The post describes several parts of a memory system: notes that capture information, a judgment about what is worth saving, and a later review process Anthropic calls “dreaming.” During that review, an agent can compare stored notes with past transcripts, correct errors and create or reorganize memory. Anthropic also describes linked subject pages, such as a wiki, and a hierarchy in which a small overview is loaded at the start of a session while other material is retrieved 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. The older model wrote 413 files across 225 folders and had earned one badge; the newer model wrote nine files in two folders plus a top-level dashboard and had earned four. Anthropic also reported that a self-built wiki raised performance on 25 company-document questions from 13% to 48% of graded criteria over 128 sessions. These figures describe the specific tests reported, not a general benchmark of agent memory.
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 Affects Agent Work
Memory determines what an agent can carry forward after a session ends. A system that saves too much can make relevant information harder to find, while one that saves too little may repeat earlier mistakes or miss a useful instruction. Anthropic’s examples point to a practical distinction: information that remains visible during a task may not need to be stored, while strategies, tested rules and ongoing plans may be useful to preserve.
The findings matter as AI products move from answering one-off prompts toward agents that work across multiple sessions. Persistent notes can help such systems retain user preferences, project details or lessons from earlier work. But a stored note can also be stale or wrong. The correction process described by Anthropic highlights that memory needs maintenance, not just storage. The company’s results offer design clues, but they do not establish that one approach will work best for every agent or task.
The article’s author also describes operating a small fleet of Claude-based agents and says that experience exposed problems with overflowing indexes, outdated notes, messages crossing between tasks and detail lost during handoffs. Those observations are the author’s account, separate from Anthropic’s evaluations; the source says Anthropic did not review or sponsor the article.
Products Worth Considering
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Sky Blue
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Black
Samsung Odyssey G93SC 49-Inch QD-OLED Curved Gaming Monitor
As an affiliate, we earn on qualifying purchases.
From Chat History to Saved Notes
For a conventional chat assistant, a new conversation generally does not automatically include the full contents of an earlier one. The model responds using what is available in its current context, such as the user’s messages and documents. Agent memory adds a separate store that can be read in later sessions, giving the model access to selected information from past work.
Anthropic’s post draws an analogy with human memory, where some sorting and consolidation takes place away from the moment an experience occurs. In its description, an agent can record information during a task and review it later, rather than treating every message as equally useful. The post connects this kind of between-session review with the idea of “sleep-time compute,” which researchers at Letta and UC Berkeley had proposed.
The examples are not evidence that an AI system remembers as a person does. They describe software mechanisms: files, summaries, retrieval and edits to stored records. Anthropic’s terminology provides a framework for thinking about those mechanisms, while its published tests show selected examples of how memory structures performed on particular tasks.
““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 Reported Tests
The results have important limits. Anthropic says its examples come from company-run 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 badge and file counts do not show how the systems would perform across repeated trials or different games.
The supplied account also gives selected results rather than enough detail to independently assess every evaluation, including the test conditions and variation across runs. It does not establish whether the same memory designs would help other models, tasks or users. Nor does it provide a general measure of how often agents’ saved notes are incorrect, become outdated or cause harmful decisions. Anthropic’s reported performance figures should therefore be read as results from the named tests, not as proof that persistent memory reliably improves agents overall.
Further Tests of Agent Memory
The immediate next step is for researchers and developers to test these approaches across more tasks and repeated runs, with clear comparisons of accuracy, retrieval quality and the cost of writing and maintaining memories. Evaluations can also examine whether agents notice when a saved rule no longer applies and whether they can trace a retrieved claim back to its source.
Anthropic’s post lays out components and examples, but the supplied material does not announce a product release or a date for further research. For now, the open question is whether structured notes and later correction will produce dependable improvements beyond the specific company tests described. The author’s account of practical problems—especially stale information and lost detail during handoffs—also points to operational issues that future evaluations will need to measure.
Key Questions
What does Anthropic mean by AI agent memory?
Anthropic describes it as information retained between sessions that a model can read later. It can include notes from user interactions or assigned tasks.
How is agent memory different from a context window?
A context window holds information available during the current session. Agent memory is saved separately so selected information can be retrieved in a later session.
What did the Pokémon test find?
In one run of each model at about 10,000 steps, Claude Sonnet 3.7 wrote 413 memory files across 225 folders and had earned one badge. Claude Fable 5 wrote nine files in two folders plus a top-level dashboard and had earned four badges. The comparison is limited to those reported runs.
Did Anthropic prove that fewer memory files are always better?
No. The post reports examples in which a smaller, more organized store was associated with better results, but it does not establish a universal rule. The evaluations were company-run, and several were single runs.
What remains uncertain about AI agent memory?
It is not clear how consistently the reported benefits hold across models and tasks, how often saved notes become inaccurate, or which maintenance methods work best. The supplied material also does not announce a timeline for further tests or a related product release.
Source: ThorstenMeyerAI.com
