🔍 Read the full analysis: AI Agent Memory In Context: Anthropic’s Research And Our Experience 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 Nov. 6, 2026, describing how Claude agents can save, organize and correct information between sessions. Its reported evaluations suggest that compact, structured notes and checks against original records can help, but the results are Anthropic’s own and include single-run comparisons. A separate account of operating Claude agents describes practical memory problems, including stale notes and lost handoff details.
Anthropic published a research post on Nov. 6, 2026, describing how its Claude models can save and organize information between conversations, then revisit those records to correct or update them. The post reports results from the company’s own evaluations, including examples in which more structured memory was associated with better task performance; the findings are not peer-reviewed, and some comparisons were single runs.
Anthropic defines agent memory as knowledge a model gains through interactions or assigned tasks and retains between sessions in a form it can read later. That differs from a model’s context window, which contains information available during the current session, and from changes to model weights during training. It also differs from summaries that replace earlier turns inside an ongoing task.
The post describes several components: memory notes, a process for judging what is worth saving, and a between-session review Anthropic calls “dreaming.” During that review, an agent can compare its notes with session logs or transcripts, correct errors and build or revise memory structures. Those structures can include a wiki of linked topic pages. A hierarchy can place a brief index at the top, while keeping detailed notes available for search when needed.
In a Pokémon Red evaluation, Anthropic compared one run each 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 created nine files in two folders, plus a top-level dashboard, and earned four badges. Anthropic argues that recording information already visible on screen can crowd out useful plans and tested rules. The comparison is illustrative, not a controlled basis for a broad performance claim.
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 Agent Memory Design Matters
Memory determines whether an agent can use lessons from one session in another, which matters when work spans multiple conversations or tasks. But the examples also show why simply saving more information is not enough: a crowded store can make useful details harder to find, while an incorrect note can steer later decisions in the wrong direction.
Anthropic’s proposed review step points to a practical trade-off for developers: agents need a way to retain useful knowledge without treating every first impression as fact. The source article also reports operational problems from its own Claude-agent use, including stale memories, full indexes, crossed messages and incomplete handoffs. These are the author’s observations, not results from Anthropic’s evaluation.
Popular right now
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Eucalyptus Green
Roborock RockNeo Q110H Wire-Free Robot Lawn Mower
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
The source article, published by Thorsten Meyer on Nov. 8, says the author reviewed Anthropic’s post and operates a small fleet of agents using Claude. It says Anthropic did not review or sponsor that account. The distinction matters: Anthropic’s examples are company-reported evaluations, while the account’s five operational lessons come from the author’s experience.
Anthropic’s post also discusses results on a database task from Continual Learning Bench 1.0, where each question began in a fresh session and a memory folder was carried forward. The post says some models produced untested guesses, while later models in its comparison included rules, evidence and warnings. It also reports a company-document task in which a self-built wiki improved the share of graded criteria met from 13% to 48% over 128 sessions. These are reported evaluation results, not independent confirmation of performance in general use.
““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 Evaluations
The results come from Anthropic’s own evaluations, presented in a research post rather than a peer-reviewed paper. The Pokémon comparison involved one run per model, so it cannot establish that the newer model’s memory approach will consistently outperform the other system. The source material does not provide enough detail to assess how broadly the reported results generalize across tasks, users or agent setups.
It is also unclear how reliably agents can identify which notes are durable, how often “dreaming” catches rather than introduces errors, or what safeguards should govern the use of sensitive information in stored memory. The article’s account of operational lessons does not provide a formal test design or quantified outcomes. Those points remain open in the material available here.
Testing Memory in More Settings
The next useful test is whether these methods hold up across repeated runs, different task types and independent evaluations, with clear reporting on errors as well as successes. Developers will also need to show how agents handle outdated or conflicting notes and how users can inspect, correct or remove stored information.
Anthropic’s post lays out a framework and examples, but the source material does not announce a product release, deployment schedule or further research milestone. For now, the practical takeaway is to treat agent memory as a system that needs organization and review—not as a guarantee that an assistant will remember correctly.
Key Questions
What did Anthropic publish?
Anthropic published a research post on Nov. 6, 2026, describing components of Claude agent memory, including saved notes, selection of useful information, between-session review and structured memory stores.
How is agent memory different from a context window?
A context window holds information the model can use in its current session. Agent memory is information saved between sessions and read later.
What does Anthropic mean by “dreaming”?
It is Anthropic’s name for an agent reviewing logs or transcripts between sessions to correct notes, add knowledge or revise memory structures.
Do the reported results prove that smaller memory stores work better?
No. The Pokémon example compared one run per model, and the post’s evaluations are company-reported rather than peer-reviewed. They illustrate Anthropic’s approach but do not establish a general rule.
What problems did the source article report from its own agents?
The author described indexes filling up, stale notes, crossed messages and loss of detail during handoffs. These are reported observations from the author’s Claude-agent use, not Anthropic evaluation findings.
Source: ThorstenMeyerAI.com
