🔍 Read the full analysis: Anthropic’s Memory Research: What It Means For AI Agents on ThorstenMeyerAI.com
Get the little things that make your day 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 research on November 6 describing how Claude agents can save and retrieve information between sessions, including memory notes, salience, reflection and linked structures. Its reported tests suggest that concise, organized notes can help on some tasks, but the results come from Anthropic’s own evaluations, include single runs and do not establish how the methods perform broadly.
Anthropic published a research post on November 6 describing how its Claude models can save, organize and correct information between sessions, a capability that can shape what an AI agent remembers when it takes up a task again. The post outlines several components of agent memory and reports results from evaluations, but the findings are Anthropic’s own tests, not a peer-reviewed study.
Anthropic distinguishes an agent’s working context—the messages and information available in the current session—from memory saved for later sessions. The company defines agent memory as knowledge acquired through interactions or assigned tasks and retained in a form the model can read later. This differs from changing the model’s trained weights, learning within a single conversation, or summarizing earlier turns to make room in the same task.
The post describes memory notes, a process for judging what is worth keeping, and a between-session reflection step Anthropic calls “dreaming.” In that step, an agent can compare saved notes with logs or transcripts, correct errors, and create or revise memory representations. The post also describes structures such as a wiki and a memory hierarchy: a short top-level file can be loaded automatically, while other material is searched or opened only when needed.
In a Pokémon Red evaluation, Anthropic compared one run each of Claude Sonnet 3.7 and Claude Fable 5 at about 10,000 steps. Sonnet 3.7 had created 413 files in 225 folders and earned one badge; Fable 5 had nine files in two folders plus a top-level dashboard and had earned four badges. Anthropic also reports that a wiki built for a company’s documents raised performance on 25 questions from 13% to 48% of graded criteria over 128 sessions. These are task-specific results, and the post notes that several evaluations involved single runs.
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 Persistent Notes Change Agent Tasks
Memory can let an agent carry forward task-specific knowledge instead of starting each session without the prior conversation. That may help with recurring work, such as applying a user’s preferences, following a project plan or reusing a verified procedure. It also makes what gets stored part of the system’s behavior: a useful rule can prevent repeated mistakes, while a stale or incorrect note can steer later work in the wrong direction.
The examples in Anthropic’s post point to a practical trade-off. Saving every detail can clutter retrieval and consume time; saving too little can discard the plan or rule that matters. A compact index and a way to inspect supporting records may help, but the published examples do not establish a general best design. For people and organizations using agents, memory quality is therefore as relevant as the model’s ability to answer within one session.
The development also has implications for oversight. If an agent relies on notes written earlier, users may need ways to inspect, correct or delete them. The post’s emphasis on checking notes against transcripts addresses one source of error, but it does not by itself settle questions about privacy, retention or how reliably corrections carry forward.
Products Worth Considering
Blueair Blue Pure 211i Max Air Purifier for Large Rooms
TheraGun Therabody Relief Massage Gun – 3 Speed, Black
Hyperice Hypervolt 3 Percussion Massage Gun with Heated Head and Carry Case
As an affiliate, we earn on qualifying purchases.
How Anthropic Defines Agent Memory
Anthropic’s post draws an analogy between an agent’s memory and the division between immediate attention and information retained over time. A model’s context window is the information it can use in the current interaction; saved notes are intended to persist beyond it. The distinction matters because an assistant that appears to have forgotten a previous instruction may have no saved record available to consult.
The post also builds on the idea of sleep-time compute, previously proposed by researchers at Letta and the University of California, Berkeley. In Anthropic’s account, an agent can do additional work between active sessions: reviewing experience, deciding what may be useful again, and organizing it for later retrieval. The post presents this as a design approach, not a claim that agent memory works like human memory in every respect.
““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 its research post
Limits of the Published Evaluations
The reported results do not show how consistently these methods work across models, tasks or users. Anthropic says several evaluations were single runs, including the comparison between Sonnet 3.7 and Fable 5 in Pokémon Red. The database and document tests measure particular tasks; their results cannot establish performance on other kinds of long-running work.
The research post is not a peer-reviewed paper, and the supplied material does not provide enough detail to independently assess every test or reproduce the reported outcomes. It is also unclear how memory quality changes over longer periods, how often corrections fail, or what safeguards and controls will be available to users. The findings support a description of Anthropic’s approach, not a general conclusion that one memory design is best.
Testing Memory Beyond These Examples
The next useful evidence would be repeated, comparable evaluations across a wider range of tasks, with methods and results detailed enough for outside researchers to assess. Those tests could show whether concise notes and between-session review improve reliability over time, and how the systems behave when saved information is outdated or contradictory.
Anthropic’s post describes the components and reports selected results, but the material here does not announce a release date or a specific product change. For now, the practical questions are how users can inspect and manage saved notes, and whether the proposed correction process reduces errors in routine use. Those points remain open beyond the research examples.
Key Questions
What did Anthropic announce?
Anthropic published a research post on November 6, 2026, describing components for Claude agent memory, including saved notes, selection of relevant information, between-session review and organized memory structures.
How is agent memory different from a context window?
The context window contains information the model can use during a current session. Agent memory refers to information saved between sessions so the model can consult it later.
What does Anthropic mean by “dreaming”?
It is Anthropic’s term for a process in which an agent reviews past logs or transcripts between sessions to correct or reorganize its notes and create new memory representations.
Do the reported results prove that memory improves all AI agents?
No. The figures come from Anthropic’s own task-specific evaluations, and several tests involved single runs. They do not establish that the approach works equally well across models or tasks.
Source: ThorstenMeyerAI.com
