🔍 Read the full analysis: AI Agents With Recall: What Anthropic’s Research And Our Runs Suggest 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 saved notes, retrieval structures and a process for correcting memories between sessions. Its reported evaluations suggest that organized, selective notes can help on some tasks, while a separate operator’s experience highlights risks such as stale or incomplete records. The findings are preliminary: several results were single runs, and Anthropic’s post was not peer-reviewed.
Anthropic published a research post on Nov. 6 describing how its Claude models can retain and organize information between sessions, a capability that can shape what an agent knows when a new conversation begins. The post reports tests in which memory organization varied alongside task performance, while a separate account from ThorstenMeyerAI.com describes practical problems with agents’ saved notes, including outdated information and lost handoff details. The reported results are early evidence, not a broad comparison: Anthropic said several evaluations were single runs, and the post was not a peer-reviewed paper.
Anthropic’s post, titled The components of agent memory, defines memory as knowledge a model gains through interactions or assigned tasks and retains between sessions in a form it can read later. That is different from a model’s context window, which holds information available during the current session, as well as from changes to model weights or summaries used to continue work within a conversation.
The company describes several parts of a memory system: memory notes that capture selected information; salience judgments about what is worth keeping; a between-session review process Anthropic calls “dreaming”; and structures such as linked wiki pages that make stored information easier to retrieve. A small top layer can be loaded at the start of each session, while other material is searched or read only when needed.
In one Pokémon Red comparison reported by Anthropic, the company examined one run each of Claude Sonnet 3.7 and Claude Fable 5 at about 10,000 game steps. Sonnet 3.7 had written 413 memory files across 225 folders and earned one badge. Fable 5 had written nine files in two folders plus a top-level dashboard file, and earned four badges. Anthropic said the dashboard included a current goal, plan and memory-use rules. The company also reported a document-question task in which a self-built wiki raised performance from 13% to 48% of graded criteria over 128 sessions; the source does not provide a broader benchmark comparison for that result.
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 can determine whether an agent carries forward a useful instruction, repeats a failed approach or acts on an outdated assumption. The research shifts attention from simply storing more conversation history to deciding what to retain, how to find it and when to correct it. That distinction matters as agents are used for tasks spanning multiple sessions, where the next session may depend on plans or rules established earlier.
Anthropic’s examples point to a trade-off: a large archive can be cumbersome, while a short, structured record may make goals and tested lessons easier to recover. But memory is not automatically reliable. ThorstenMeyerAI.com, which says it operates a small fleet of Claude-based agents and was not sponsored or reviewed by Anthropic, reports issues including a crowded index, stale notes, crossed messages and incomplete handoffs. Those observations are an operator’s account, not a controlled evaluation, but they underline that persistent memory can carry errors forward as well as useful knowledge.
For users and developers, the practical question is not just whether an agent remembers, but whether its notes are accurate, relevant and traceable. A saved rule without its date, source or reason may be misleading later. The systems described remain tools whose outputs and actions need appropriate checks, especially when errors could have real consequences.
Products Worth Considering
Apple MacBook Pro 14.2-inch with M5 Max, 18-Core CPU, 20-Core GPU, 48GB Memory, 2TB SSD, Space Black
Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID (Midnight)
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Sky Blue
As an affiliate, we earn on qualifying purchases.
From Session Context to Saved Notes
Traditional chat sessions rely on information visible in the current context window. When the session ends, that working context may not be available in the next one. Agent memory adds a separate store that can be consulted later, more like a filing system than a continuously open transcript.
Anthropic’s framework draws an analogy to human memory, where some processing happens away from immediate attention. It calls the agent’s between-session review “dreaming”: reviewing logs or transcripts to revise notes, add representations or decide what should be loaded next time. Anthropic connects the idea to “sleep-time compute,” which it says was proposed by researchers at Letta and UC Berkeley. The post presents a system design and evaluations, rather than evidence that AI memory works in the same way as human memory.
The source article also describes lessons from its own Claude-based agents: indexes can fill, old notes can become unsafe to rely on, messages can be mixed up, and handoffs can omit details. Its recommendation to write rules with a date and reason is an operational lesson from that experience, not a finding independently established by Anthropic’s tests.
“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 Reported Memory Tests
The scale and repeatability of the results remain unclear. Anthropic describes several evaluations as single runs, including the Pokémon comparison, so the reported outcomes do not establish that one memory design will consistently outperform another. The post is a research post rather than a peer-reviewed paper, and the provided material does not include enough detail to independently assess every test or reproduce the results.
The reported 13% to 48% result is described as a share of graded criteria across 128 sessions, but the source does not specify a comparison baseline beyond the before-and-after figures or establish how well that task represents other uses. It is also not clear how the memory systems perform over longer periods, when notes become outdated, or when several users or tasks may produce conflicting information. The account from ThorstenMeyerAI.com offers practical observations but does not provide controlled measurements or a comparison group.
More Testing of Persistent Agent Memory
Anthropic’s post sets out components and examples, but the material provided does not announce a product launch, deployment schedule or follow-up study. Further evidence would need to show whether the reported benefits hold across repeated runs, different tasks and longer periods of use, and how systems detect and correct stale or conflicting notes.
For now, developers and users assessing memory-enabled agents will need to check what is saved, what is retrieved and whether stored instructions reflect current information. The next meaningful milestone is broader, reproducible testing that reports both task performance and memory errors, rather than file counts alone.
Key Questions
What did Anthropic announce?
Anthropic published a research post on Nov. 6, 2026, describing components of AI agent memory: saved notes, decisions about what matters, between-session review and structures for finding information later. The source does not describe a new product launch.
How is agent memory different from a context window?
A context window contains information available to a model during a session. Agent memory is information saved between sessions and read later. Anthropic’s definition also separates memory from model-weight changes and summaries used to continue work within a session.
Did Anthropic prove that fewer memory files work better?
No. Anthropic reported one Pokémon Red run each for Claude Sonnet 3.7 and Claude Fable 5, with different file counts and badge totals. Those results suggest a possible relationship between memory organization and task performance, but single runs do not establish a general rule.
What problems can persistent agent memory create?
Saved notes can become outdated, incomplete or difficult to retrieve, and information from separate messages or tasks can be mixed up. ThorstenMeyerAI.com reports these issues from its own agent operations; that account is not a controlled study.
What remains to be tested?
Further work would need to establish how repeatable the reported gains are across tasks and models, how systems identify stale or conflicting notes, and whether memory remains reliable over longer periods. The source material does not give a schedule for additional Anthropic results.
Source: ThorstenMeyerAI.com
