🔍 Read the full analysis: The Research Behind AI Agent Memory—and Our Experience Running It on ThorstenMeyerAI.com
Get the latest gadgets 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 describing how Claude agents can save, organize, retrieve and correct information across sessions. Its reported tests suggest that selective, structured notes can help on some tasks, while the post and ThorstenMeyerAI.com’s own experience also point to risks from stale or poorly managed memory.
Anthropic published a research post on November 6, 2026, describing how its Claude models can save, organize and correct information between conversations—a capability that can affect what an AI agent remembers when it handles a later task. The post reports results from company-run evaluations, while ThorstenMeyerAI.com says its own small fleet of Claude agents has also exposed practical problems, including stale notes and lost handoff details.
Anthropic defines agent memory as knowledge a model gains from interactions or assigned tasks and retains between sessions in material it can read later. That is separate from a model’s context window, which contains information available during a current session, and from changes to model weights made during training. It is also distinct from summaries used to continue work after a context window fills.
The post describes several memory components: notes written during or after a task; salience decisions about what is likely to matter again; a between-session review Anthropic calls “dreaming”; and organized representations, such as linked wiki pages. It also outlines a memory hierarchy, with a small set of top-level notes available at the start of each session and other material retrieved only when needed.
Anthropic’s examples suggest that more stored information is not automatically better. In one Pokémon Red comparison at about 10,000 game steps, a single run with Claude Sonnet 3.7 had created 413 files across 225 folders and earned one badge. A single Claude Fable 5 run had created nine files in two folders plus a top-level dashboard file and earned four badges. The post also reports that a wiki built from company documents raised performance on 25 questions from 13% to 48% of graded criteria over 128 sessions.
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
Memory Quality Shapes Later Tasks
Persistent memory gives agents a way to carry user preferences, task rules and discoveries across sessions, potentially reducing repeated explanations and helping them pick up ongoing work. But a saved note can also mislead a later session if it is wrong, out of date or retrieved without the needed qualifications. The development matters because memory is not just a storage feature: what an agent chooses to retain can influence its subsequent decisions.
The examples point to a trade-off between keeping enough information to support future work and limiting clutter. Anthropic’s game comparison is only one run per model, so it does not establish that a particular memory format will reliably improve performance. The company-document result is more structured, but still comes from the company’s own evaluation. For organizations deploying agents, the practical questions include how memories are checked, updated, scoped to the right user or project, and removed when they no longer apply.
ThorstenMeyerAI.com reports five operational lessons from its own agents: indexes can fill up, old notes can become dangerous, messages can cross, handoffs can lose detail, and rules are more useful when recorded with a date and a reason. These are the publisher’s observations, not independently verified findings or results reviewed by Anthropic.
Products Worth Considering
Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, 12MP Center Stage Camera, and Wi-Fi 7, Sky Blue
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 24GB Memory, 1TB SSD, Midnight
Apple MacBook Air 15-inch (2025) with M4 Chip, 24GB Memory, 512GB SSD, Midnight
As an affiliate, we earn on qualifying purchases.
From Blank Sessions to Saved Notes
Many chat assistants historically began each new conversation without access to the previous one. A context window can hold a conversation and supplied documents while a session is active, but that material does not by itself become durable memory. Agent systems can instead save selected information externally and load or search it during later work.
Anthropic’s post draws an analogy to how human memory is consolidated during periods of rest, describing an agent’s between-session review as “dreaming.” The idea builds on the term “sleep-time compute,” which the source says was proposed by researchers at Letta and UC Berkeley. In Anthropic’s account, this review can compare notes with original logs, correct errors, form new representations and decide what to load later.
The reported evidence has important limits: the post is company-published research, not a peer-reviewed paper, and some comparisons are single runs. The source says Claude Fable 5 and other model versions were evaluated, but the results should be read as Anthropic’s reported findings rather than a general, independent ranking of agent memory systems.
“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 Tests
Anthropic’s published results do not establish how consistently the memory approaches work across models, tasks or production settings. The Pokémon comparison involved one run per model, and the source identifies several other results as company evaluations rather than peer-reviewed research. It is also unclear from the supplied material how the reported gains would compare with alternative memory designs under the same independent test conditions.
The source does not give a full technical account of how memory errors, privacy, access boundaries or deletion are handled in every deployment. Those issues matter when notes include personal or sensitive information, or when several agents and users share systems. ThorstenMeyerAI.com’s account describes practical failure modes but does not provide underlying measurements or a controlled comparison. Anthropic did not review or sponsor that account, according to the publisher.
Testing Memory in Real Deployments
The next meaningful evidence will come from repeated tests across a wider range of tasks and from deployments that report how often agents retrieve relevant notes, rely on stale information or correct mistakes. Readers should watch for independent evaluations that compare memory designs using shared tasks and clear performance measures, rather than treating individual examples as definitive.
For teams already using agents, the immediate work is operational: review what gets saved, make time-sensitive rules explicit, check that notes remain tied to the right project, and retain enough source history to correct errors. Anthropic’s post describes a method for thinking about these systems; the available material does not set out a new product launch or a confirmed release timetable. Further results and deployment details remain to be seen.
Key Questions
What did Anthropic publish?
Anthropic published a research post titled “The components of agent memory” on November 6, 2026. It describes how Claude agents can record, organize, retrieve and revise information between sessions.
How is agent memory different from a context window?
A context window holds information the model can use during a current session. Agent memory consists of selected information saved outside that session and made available to the agent later.
Do Anthropic’s examples prove that structured memory improves agents?
No. The examples are Anthropic-reported evaluations, and the Pokémon comparison used one run per model. They offer evidence about the tested cases but do not establish performance across all tasks or deployments.
What problems can persistent agent memory create?
Saved notes can become outdated, contain incorrect assumptions, be retrieved in the wrong setting or omit details needed for a handoff. ThorstenMeyerAI.com reports encountering these issues with its own Claude-based agents.
What does Anthropic mean by “dreaming”?
Anthropic uses “dreaming” for a between-session review in which an agent checks logs and transcripts, corrects or adds notes, and prepares information for future sessions.
Source: ThorstenMeyerAI.com
