🔍 Read the full analysis: Running Agents With Memory: What AI Research Can—and Can’t—Tell Us on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Anthropic published a research post on November 6, 2026, describing how Claude agents can save, organize and revise information between sessions. Its reported tests suggest that selective, structured memory can help on particular tasks, but the results are limited and do not establish how reliably memory works in broader use.
Anthropic published a research post on November 6, 2026, describing how its Claude models can save and organize information between conversations, then revisit it to correct or update their notes. The post reports better results from more selective, structured memory in some company-run tests, but it is not a peer-reviewed paper, and the examples do not establish how well the approach will work across everyday tasks or other agent systems.
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 differs from a model’s context window, which holds information available during a current session. It also differs from changing model weights, learning within one conversation, or summarizing earlier turns to make room in a context window.
The post describes several components: memory notes that capture potentially useful information; a salience process for deciding what merits saving; “dreaming,” in which an agent reviews earlier logs between sessions and revises its notes; and organized representations, such as linked wiki pages. Anthropic also describes a hierarchy in which a short overview is loaded at the start of a session, while more detailed material is retrieved when needed.
In one Pokémon Red comparison, Anthropic reported that a single run of Claude Sonnet 3.7 created 413 memory files across 225 folders and had earned one badge at about 10,000 game steps. A single Claude Fable 5 run created nine files in two folders plus a top-level dashboard file, and had earned four badges. The post says the newer model’s compact notes held plans and tested rules, while much of the older model’s material duplicated information visible on screen. These are results from individual runs, not a controlled estimate of typical performance.
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 Affects Agent Work
Persistent memory is one way to address a basic limitation of session-based assistants: a system may not have access to what a user said in an earlier conversation. Saving notes can support continuity, but the notes also shape what the agent believes and does later. Memory quality, not just memory volume, therefore matters: a concise, tested rule may be more useful than a large archive of unverified details.
The issue is practical as well as technical. If a stored deadline is outdated, an agent could plan around the wrong date; if a handoff leaves out a constraint, later work may miss it. The source publisher describes its own small fleet of Claude agents as encountering problems with full indexes, stale notes, messages crossing and lost handoff details. Those are the publisher’s operational observations, not findings independently evaluated by Anthropic or a broad study.
Anthropic’s examples offer a framework for thinking about memory systems, not a guarantee of dependable recall. They also point to a trade-off: saving less can reduce clutter, but an agent must still identify what matters, preserve evidence for it and retrieve it at the right time.
Products Worth Considering
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Sky Blue
Logitech MX Creative Console with 9 Customizable Keys, Graphite
Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite
As an affiliate, we earn on qualifying purchases.
How Anthropic Frames Agent Memory
Anthropic’s post draws an analogy with human memory, where some sorting and consolidation occurs away from immediate experience. In the proposed agent version, the model records notes during work and later reviews transcripts to correct mistakes, add useful information or change how future sessions are supplied with context. Anthropic calls that later review step “dreaming.”
The idea builds on “sleep-time compute,” which the source says was proposed by researchers at Letta and UC Berkeley. The memory materials described can range from a folder of notes to a linked wiki, with a brief index at the top and details accessed as needed. The source also reports a company-document test in which a self-built wiki raised scores across 25 questions from 13% to 48% of graded criteria over 128 sessions. That figure is the reported evaluation result; it does not by itself establish performance on other document collections.
The account and the examples here come from a ThorstenMeyerAI.com report dated November 8, which says its author runs a small fleet of Claude-based agents and that Anthropic did not review or sponsor the article. Anthropic’s research post, as described in that report, is the primary basis for the model evaluations discussed here.
““Dreaming””
— Anthropic, as characterized in its research post
Limits of the Reported Tests
The evidence described is limited to Anthropic’s own evaluations, published in a research post rather than a peer-reviewed paper. The Pokémon comparison involved one run per model, so the badge counts and file totals do not show how often the same outcome would occur in repeated trials. The source also does not provide enough detail here to assess statistical uncertainty or independently reproduce the reported results.
It remains unclear how often agents will save inaccurate or outdated information, how consistently their later review will catch it, and how well the methods transfer to different models, users and work settings. The source’s account of memory problems in its own agent fleet is anecdotal. It does not establish rates of failure, and the article does not report an independent audit of those systems.
Further Tests Needed Across Tasks
The next evidence readers should look for is repeated, independently reported testing across a wider set of tasks and systems. Such evaluations could compare memory methods over multiple runs, measure whether saved facts remain accurate over time and report how much retrieval or correction changes task outcomes. The available source material does not identify a date for a follow-up study or a separate next milestone from Anthropic.
For now, the research post offers a description of how Anthropic approaches agent memory and examples of results in selected evaluations. Whether the same designs improve reliability in routine use remains an open question.
Key Questions
What did Anthropic publish?
Anthropic published a research post on November 6, 2026, describing components for Claude agent memory, including saved notes, salience, later review called “dreaming,” and organized memory structures.
Is agent memory the same as a context window?
No. A context window holds information available during a session. Agent memory refers to information saved between sessions and made available for later use.
Do the reported results prove that more memory improves agents?
No. The examples suggest that selective, structured memory helped in some Anthropic evaluations. The Pokémon comparison included one run per model, and the reported tests do not establish general performance across tasks or systems.
What does Anthropic mean by “dreaming”?
It is Anthropic’s name for a between-session process in which an agent reviews earlier logs or transcripts and may correct, add or reorganize stored memory.
What remains uncertain about AI agent memory?
The available account does not establish how reliably agents detect stale or incorrect notes, how often the approach works in repeated trials, or whether the results transfer to other models and real-world settings.
Source: ThorstenMeyerAI.com
