🔍 Read the full analysis: How To Think About AI Agent Memory: Research Meets Practice 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 a research post on Nov. 6 describing how Claude agents can save information between sessions, organize it for later use and revise errors. Its reported evaluations suggest that concise, structured notes can outperform sprawling memory stores, but the results include single runs and are not peer-reviewed.
Anthropic published a research post on November 6, 2026, describing how AI agents can retain and revise information between conversations, rather than starting each session without prior notes. The post lays out memory components including note-taking, decisions about what matters, later review and layered storage, and reports results from evaluations of Claude models; those results are Anthropic’s own findings, not peer-reviewed research.
Anthropic distinguishes an agent’s context window—the information available during a current session—from memory saved for later sessions. Its definition covers knowledge gained through interactions and tasks that is retained in a form the model can read again. It excludes changes to model weights, learning within one conversation and summaries used to make room in an ongoing session.
The post describes several parts of a memory system. An agent creates memory notes and makes a judgment about which experiences are likely to matter again. In a later, offline step Anthropic calls “dreaming,” it can compare notes with earlier transcripts, correct errors or create new representations. These may include a linked wiki, with subject-specific pages, and a memory hierarchy that loads a small overview automatically while keeping other information available for search.
In one Pokémon Red evaluation, Anthropic compared single runs of Claude Sonnet 3.7 and Claude Fable 5 at about 10,000 game steps. Sonnet 3.7 had created 413 files across 225 folders and earned one badge; Fable 5 had nine files in two folders, plus a top-level dashboard, and earned four badges. Anthropic says visible game information did not need to be saved repeatedly, while plans and tested strategies were more useful. The comparison is a single-run result, not evidence that the newer model will always perform 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 Work
Memory design can shape whether an agent follows a user’s latest instruction, reuses a verified procedure or repeats a past mistake. A larger archive is not automatically more useful: irrelevant, duplicated or outdated notes can crowd out the information needed for a task. Anthropic’s examples make selection and correction central design questions, alongside storage capacity.
The distinction matters for people and organizations using agents across multiple sessions. A system that carries forward a deadline or project decision may save repeated explanation, but only if it records the information accurately and makes it available at the right time. If memory is treated as unquestioned truth, an old assumption can continue affecting later work. The post’s proposed review step offers one approach to that risk, though it does not establish how reliably it works across real-world tasks.
Anthropic also reports a company-document task in which a self-built wiki increased performance from 13% to 48% of graded criteria across 128 sessions and 25 questions. That result suggests structured references may help on questions requiring information gathered over time. It does not, on its own, establish that wikis improve every agent task or that the same gain would occur in other settings.
Products Worth Considering
Levoit Core 400S-P Smart Air Purifier for Large Rooms
LEVOIT Vital 200S-P Smart Air Purifier for Large Rooms
Shark HydroVac Reveal Cordless Wet-Dry Vacuum and Mop, Lilac, WD321
As an affiliate, we earn on qualifying purchases.
From Session Context to Saved Notes
Many chat systems have historically treated each conversation as separate: the model can use what is visible in the current context but may not carry it into the next session. Agent memory adds a store that the system can consult later. It is distinct from the model’s training and from a summary that merely helps an ongoing conversation fit within a context window.
Anthropic’s post draws an analogy to human memory, arguing that some sorting and consolidation can happen outside the immediate task. It also connects its “dreaming” idea to sleep-time compute, a concept proposed by researchers at Letta and the University of California, Berkeley. The article describes the analogy as a design approach for agents, not as evidence that machine memory works like human memory.
The source article also reports practical lessons from its author’s own small fleet of Claude-based agents. These include risks from crowded indexes, stale notes, messages crossing between tasks and handoffs losing detail. The author says Anthropic did not review or sponsor that account. Those observations are operational experience, separate from Anthropic’s published 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, defining agent memory in its research post
Limits of the Published Evaluations
The findings are presented in an Anthropic research post, not a peer-reviewed paper. The source says several results are single runs, including the Pokémon comparison, which makes it difficult to know how repeatable the outcomes are. The reported tasks also do not show how the methods perform across the range of agents, users and high-stakes settings.
The available material does not establish how often an agent’s “dreaming” step detects a false note, how much review costs in time or computing resources, or how systems should handle conflicting instructions and sensitive information. The source article’s author describes practical problems from a small set of agents, but does not provide a formal study or sample size. Those observations should not be treated as general performance measurements.
Testing Memory Beyond Examples
The next useful evidence would include repeated trials, comparisons across different tasks and models, and measurements of how often saved notes are accurate, relevant and corrected when wrong. Costs and failure rates would also help teams decide whether automatic note-writing and later review are worth using.
For now, Anthropic’s post offers a description of its approach and a set of reported evaluation results. Readers assessing an agent should distinguish what it can see in the current session from what it stores for later, and ask how those stored notes are checked. Further evaluation is needed before the examples can establish dependable results beyond the tasks reported.
Key Questions
What does Anthropic mean by AI agent memory?
It means information an agent gains from interactions or assigned tasks and saves in a form it can read in later sessions. Anthropic distinguishes this from the current conversation’s context window and from changes to model weights.
What is “dreaming” in Anthropic’s framework?
Anthropic uses “dreaming” for a process between sessions in which an agent reviews logs and transcripts, checks existing notes and may correct or reorganize them. The post presents it as a memory-system component, not a claim that an AI sleeps like a person.
Did the Pokémon test prove that smaller memory stores are better?
No. Anthropic reported one run each of two Claude models. The newer model used fewer files and earned more badges in that comparison, but a single-run result cannot establish that smaller memory stores will perform better in general.
How strong is the evidence in the post?
The findings come from Anthropic’s own evaluations, published as a research post. The source says several results are single runs, and the post is not described as peer-reviewed research.
Source: ThorstenMeyerAI.com
