🔍 Read the full analysis: What We Learned About AI Memory From Anthropic’s 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’s November 6, 2026 research post describes how Claude agents can save information between sessions, correct notes and organize them for later use. Its published evaluations suggest that structured, selective memory can outperform sprawling note collections, though several results come from single runs and Anthropic’s own tests.
Anthropic published a research post on November 6 describing how its Claude models can save, organize and revise information between conversations, a capability that can affect what an AI agent remembers when it starts a new task. The post outlines memory notes, selection, review between sessions and layered ways to retrieve stored information, and reports evaluation results suggesting that organized notes can be more useful than simply saving more. Anthropic presented the findings as a research post, not a peer-reviewed paper.
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. This differs from a model’s context window, which contains information available during the current session, and from changes to model weights during training. It also distinguishes memory from summaries that replace earlier conversation turns within an ongoing task.
The post describes several components. An agent writes memory notes about information it judges useful later; salience is the process of deciding what merits saving. Anthropic calls a between-session review of logs and transcripts “dreaming,” during which an agent can check or revise notes and create new representations. These representations can include a wiki of linked subject pages. A hierarchy can load a small set of top-level notes automatically, while leaving other material to be searched when needed.
Anthropic reports a Pokémon Red comparison involving one run each of Claude Sonnet 3.7 and Claude Fable 5, at about 10,000 steps into the game. Sonnet 3.7 had created 413 files in 225 folders and earned one badge. Fable 5 had created nine files in two folders, plus a top-level dashboard, and earned four badges. Anthropic’s interpretation is that repeated information visible in the game need not be stored, while plans and tested strategies may be worth keeping. The result is a comparison of those runs, not a general benchmark proving one memory design always performs better.
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 make an agent’s behavior across sessions depend on more than its current prompt: it also depends on what the system chose to retain and whether that information remains accurate. A useful note might preserve a tested procedure or a user’s stated preference; a poor one might retain a mistaken guess, omit a key qualification or direct the agent toward an outdated plan.
That makes memory design relevant to people using agents for work that continues over multiple conversations. A compact index, searchable supporting notes and a way to revisit original records could help an agent retrieve useful information without loading an entire history. Anthropic’s examples also show why raw storage volume is a weak measure of memory quality: its Pokémon test favored a smaller set of notes in one comparison, while another reported evaluation examined whether an agent’s stored rules had been checked.
Those findings are evidence from Anthropic’s own evaluations, not independent confirmation that the methods will improve every model or task. The post helps describe a technical design problem and possible approaches; it does not establish that persistent agent memory is reliably accurate in ordinary use.
Products Worth Considering
Apple MacBook Air 13.6-inch with M5 chip, 16GB Memory, 512GB SSD, Midnight
Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Midnight
Apple MacBook Neo 13-inch Laptop with A18 Pro Chip, 8GB Memory, 256GB SSD, Liquid Retina Display, and 1080p Camera – Blush
As an affiliate, we earn on qualifying purchases.
From Session Context to Saved Notes
Many chat systems answer from information present in the current conversation. When a session ends, that working context may not carry into the next one. Anthropic’s post addresses a separate layer: information an agent saves so it can consult it later. The company’s definition excludes both the model’s trained parameters and short-term summaries used to continue a task after a context window fills.
The research post also draws on the idea that memory can be processed between active tasks. Anthropic compares this to the way human memories are thought to be sorted during sleep and rest, while describing an AI-specific mechanism for reviewing transcripts. Its use of “dreaming” refers to that proposed agent process, not to a claim that a model sleeps or has human experiences. The post builds on the concept of sleep-time compute, which it says was proposed by researchers at Letta and UC Berkeley.
““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’s research post, “The components of agent memory”
Limits of Anthropic’s Evaluation Results
The research post reports company-run evaluations, and the supplied account says several results are single runs. The Pokémon comparison involved one run per model, so it cannot establish how consistently either model would perform under repeated trials or different game conditions. The reported badge counts and file totals are observations from that test, not a universal ranking of memory systems.
The material also describes a company-document question test in which a self-built wiki raised scores from 13% to 48% of graded criteria over 128 sessions. It does not provide enough detail here to judge the full evaluation design, scoring procedure or how broadly the result applies. The account begins to describe results from a database task on Continual Learning Bench but does not include all findings, so a complete comparison of the models in that test cannot be made from the available material.
It remains unclear how often saved notes become stale in deployment, how well the correction process catches errors, and what trade-offs memory systems create for privacy, cost and response speed. Anthropic’s examples illustrate possible methods; they do not settle those questions.
Testing Memory Beyond Published Examples
The next evidence to watch for is whether Anthropic or independent researchers publish repeatable evaluations across more tasks, models and runs. Useful comparisons would test not only task scores but also how often agents retrieve relevant notes, preserve accurate details, correct errors and avoid relying on outdated information.
Anthropic’s post describes a design in which agents can check their notes against source transcripts between sessions and load only selected memory layers when needed. Further reporting and testing would be needed to establish how those systems perform in routine use, and what controls users have over what is stored. No additional release or follow-up date is specified in the supplied source material.
Key Questions
What did Anthropic publish?
Anthropic published “The components of agent memory” on November 6, 2026. The research post describes how Claude agents can save information between sessions, review it and organize it for later retrieval.
How is agent memory different from a context window?
A context window holds information available to a model during a session. Agent memory refers to information saved between sessions so the model can read it later.
Did the Pokémon test prove that fewer memory files are better?
No. Anthropic reported one run each for two models. In that comparison, Claude Fable 5 used fewer files and earned more badges than Claude Sonnet 3.7, but the test does not establish a general rule across models or tasks.
What does Anthropic mean by “dreaming”?
It is Anthropic’s term for an agent reviewing logs or transcripts between sessions to check existing notes, correct mistakes or form new memory representations. It describes a proposed computing process, not human-like sleep.
Are the findings independently verified?
The supplied material describes the results as coming from Anthropic’s own evaluations and says the post is not a peer-reviewed paper. Independent replication and broader testing are not established in the available account.
Source: ThorstenMeyerAI.com
