🔍 Read the full analysis: A Closer Look At AI Agent Memory And What We Learned Using It 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 evaluations suggest that compact, structured memory can help on some tasks, but the results are limited and include single runs; an independent operator also describes practical problems with stale or crowded notes.
Anthropic published a research post on November 6, 2026, describing how its Claude models can save and retrieve information between conversations, and how they may revise those notes over time. The post reports results from company-run evaluations, while an article published by ThorstenMeyerAI.com on November 8 describes practical memory problems observed with its own Claude-based agents; Anthropic did not review or sponsor that account.
Anthropic distinguishes an agent’s context window—the information available during the current session—from memory saved for later sessions. It describes memory as knowledge gained from interactions or assigned tasks and retained in a form the agent can read again. That category excludes changes to model weights, learning within a single conversation and summaries used to continue work in a new context window.
The post identifies several components: notes written during or after a task, a decision about which information is worth keeping, a later review process Anthropic calls “dreaming,” and structures such as linked wiki pages. A small top-level memory layer may be loaded automatically, while other files can be searched or read only when relevant. The post draws an analogy with offline memory processing during sleep, but describes an AI system design rather than evidence that models remember as people do.
In one Pokémon Red comparison at roughly 10,000 game steps, Anthropic reported that Claude Sonnet 3.7 created 413 files in 225 folders and had earned one badge. Claude Fable 5 created nine files in two folders plus a top-level dashboard and had earned four badges. Anthropic also reports that a self-built wiki improved performance on 25 questions about company documents from 13% to 48% of graded criteria over 128 sessions. The post says several results were single runs; they are company-reported evaluations, not peer-reviewed findings.
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 Design Shapes Agent Results
Persistent memory can let an agent carry forward plans, user preferences and lessons from earlier tasks instead of restarting with a blank record. That makes its design relevant to anyone using agents across multiple sessions: the saved notes may affect what the system does later, even when a user cannot see how those notes were selected or updated.
Anthropic’s examples point to a trade-off between storing detail and keeping memory useful. Its Pokémon comparison associates a smaller, organized set of notes with better progress in that particular run, while the company-document test reports higher scores with a maintained wiki. These examples do not establish that less memory or a wiki will improve performance across models and tasks. They do show why memory quality, retrieval and correction are practical parts of agent reliability, not simply a matter of increasing the context window.
The ThorstenMeyerAI.com account adds operational concerns: its author says indexes can fill up, stale notes can mislead agents, messages can become mixed up, and handoffs can lose detail. Those are observations from the author’s own agents, not results independently verified by Anthropic. Taken together, the material highlights a central risk: a system may confidently act on a retained note that is outdated or wrong unless it can check and correct it.
Popular right now
Dyson Supersonic Nural Hair Dryer
Bose QuietComfort Wireless Noise Cancelling Headphones, Twilight Blue
Vitamix Ascent X3 Professional-Grade Blender, 48 oz, Shadow Black
As an affiliate, we earn on qualifying purchases.
From Session Notes to Agent Memory
For a chat assistant without persistent memory, each new conversation begins without the previous session’s content. The current context window functions as temporary working space; saved agent memory is a separate store that can be read in a later session. Anthropic’s post focuses on that between-session process rather than on how a model’s original training changes.
The approach described combines immediate note-taking with later review. An agent can record information it judges likely to matter again, then revisit session records to correct errors or reorganize what it has learned. The post also builds on the idea of “sleep-time compute,” which it attributes to prior proposals from researchers at Letta and UC Berkeley. Anthropic presents this as a design framework, not a claim that every Claude agent uses every component in the same way.
The source article says its author runs a small fleet of Claude-based agents and learned five practical lessons during the week described. The account says dated rules, including a reason for each rule, work better for its agents. It is an operator’s report, separate from Anthropic’s evaluations, and should not be read as a controlled comparison.
““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 “The components of agent memory”
Limits of the Reported Evidence
The post’s results come from Anthropic’s own evaluations, not a peer-reviewed paper. The source says several comparisons were single runs, including the Pokémon example. The material provided does not establish whether the reported results would hold across repeated trials, different tasks, or other model families.
It is also unclear how often agents using these designs make incorrect updates, retain sensitive information, or retrieve a misleading note. The source article’s account of stale memory and crossed messages is anecdotal and does not quantify how often those problems occurred. No broader reliability or privacy figures are provided in the supplied material.
Testing Memory Across More Tasks
The next useful test is whether the reported gains persist across repeated runs and a wider range of tasks, with clear comparisons between memory structures and models. The supplied source does not announce a timetable for further Anthropic evaluations or a new product release. For now, the November 6 post offers a description of memory components and selected company-run results; the durability and generality of those results remain open.
Key Questions
What did Anthropic publish?
Anthropic published “The components of agent memory” on November 6, 2026. The research post describes how Claude agents can write, organize, retrieve and revise information saved between sessions.
How is agent memory different from a context window?
The context window contains information available during the current session. Agent memory is saved information that the system may read in a later session.
What does Anthropic mean by “dreaming”?
Anthropic uses “dreaming” for a between-session review of logs or transcripts. The agent may use that review to correct notes, add information or reorganize its memory.
Do the reported results prove that smaller memory stores are better?
No. The Pokémon result was a single-run comparison, according to the source, and the research post is not peer-reviewed. It suggests that a more compact, structured store helped in that example, but does not show that the result generalizes.
What problems did the independent account report?
ThorstenMeyerAI.com’s author said its agents encountered crowded indexes, stale notes, mixed messages and loss of detail during handoffs. These are the author’s reported observations, not independently verified findings or quantified rates.
Source: ThorstenMeyerAI.com
