🔍 Read the full analysis: AI Memory For Agents: What Research Shows And Experience Adds on ThorstenMeyerAI.com
Get business pricing on tech for your team
- 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 AI agents can save, organize and correct information between sessions. Its reported tests suggest that selective, structured notes can help on some tasks, but the results come from Anthropic’s own evaluations, including single runs. A separate account of Claude-based agents describes practical problems such as stale notes and information lost in handoffs.
Anthropic published a research post on November 6, 2026, describing how its Claude models can save and retrieve information between conversations, a capability that can shape what an agent remembers when a new task begins. The post reports that structured, selective memory performed better than a large collection of loosely organized notes in some of Anthropic’s evaluations, though the results are company-reported and several comparisons were single runs.
Anthropic’s post, “The components of agent memory,” defines memory as knowledge an agent gains from interactions or assigned tasks and retains between sessions in a form it can read later. It distinguishes that from a model’s context window—the information available during a current session—as well as from changes to model weights, learning within one conversation and summaries used to continue work in a new context window.
The system Anthropic describes has several parts: an agent writes memory notes, decides which information is worth retaining, and may later review its logs to correct or reorganize those notes. Anthropic calls this between-session review “dreaming.” It also describes memory representations such as a linked wiki and a hierarchy of notes, with a small amount loaded automatically and other material retrieved when needed.
In one 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 said the comparison illustrates the value of retaining plans and tested rules rather than recording information already visible in the game; a single comparison does not establish that the newer approach will perform better across tasks.
In a separate database task from Continual Learning Bench 1.0, each question began in a fresh session, with a memory folder as the carryover. Anthropic reported that Claude Sonnet 4.6 produced notes containing untested guesses, while Claude Opus 5 and later models in the comparison recorded rules along with evidence and warnings. Another evaluation, involving 25 questions about company documents over 128 sessions, reportedly saw graded criteria scores rise from 13% to 48% after an agent built a wiki. These are Anthropic-reported findings, not peer-reviewed results.
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 Structured Agent Memory Matters
Memory changes the way an AI assistant can handle work that spans multiple conversations. Without saved information, a new session may not know a deadline, preference or decision established earlier. With memory, an agent can carry forward useful details—but it can also carry forward mistakes, outdated instructions or unsupported guesses.
Anthropic’s examples point to a practical trade-off: saving more information is not automatically better. A concise index and well-supported notes may be easier to use than a sprawling archive. For users and developers, the quality of an agent’s memory process—what it keeps, how it checks its notes and when it retrieves them—can matter as much as the amount of information available.
A separate account accompanying the research describes five operational problems seen while running Claude-based agents: an index that fills up, stale information, messages crossing between tasks, detail lost during handoffs, and rules that work better when written down with a date and reason. The author says Anthropic did not review or sponsor that account. These are practitioner observations, not controlled research findings, but they highlight issues that evaluation scores alone may not capture.
Products Worth Considering
Apple MacBook Air 15-inch (2025) with M4 Chip, 24GB Memory, 512GB SSD, Midnight
Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, 12MP Center Stage Camera, and Wi-Fi 7, Sky Blue
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Raspberry
As an affiliate, we earn on qualifying purchases.
From Blank Sessions to Saved Notes
For a chat assistant without a memory feature, information generally stays within the active conversation. The context window is the working space the model can use at that moment; it is not, by itself, a durable record available in a later session. Agent memory adds a separate store that can be read again.
Anthropic’s proposed organization also draws a distinction between recording events and building reusable knowledge. An agent may tag important moments as they happen, then review transcripts between sessions to resolve errors, identify recurring rules or create topic-specific pages. The research post connects this review process to “sleep-time compute,” an idea proposed by researchers at Letta and UC Berkeley. The post’s neuroscience analogy is an explanatory model, not evidence that machine memory works like human memory.
The account’s reported evaluations use different tasks and model versions, so their figures should not be treated as one unified benchmark. Anthropic presents them as examples of design choices and performance on particular tests, rather than as proof that one memory architecture is best for every agent.
“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 Evaluations
The results described are from Anthropic’s own evaluations, published in a research post rather than a peer-reviewed paper. The Pokémon comparison involved one run per model, and the supplied account also characterizes several results as single runs. It is not clear from the material provided how consistently the reported differences would appear across repeated trials, other tasks or agents from different developers.
The account does not provide full details for every benchmark result, including enough information to independently assess all scoring procedures or reproduce the evaluations. It also does not establish how often memory errors occur in everyday use, or how reliably an agent can detect that a saved note has become outdated. The practical lessons about handoffs, mixed messages and stale notes are the author’s observations, not independently verified measurements.
Testing Memory Beyond These Examples
The next meaningful test is whether the reported gains hold across repeated runs, a wider range of tasks and independent evaluations. Readers should look for details on benchmark methods, error rates and how systems decide when to write, revise or retrieve notes. Those measures would help distinguish improvements from results specific to a particular setup.
For now, Anthropic’s post offers a description of its approach and examples of how memory organization can affect task performance. The accompanying practitioner account adds questions about day-to-day reliability. Whether these methods generalize, and how well agents handle outdated or incorrect notes in routine use, remains unresolved in the supplied material.
Key Questions
What did Anthropic announce?
Anthropic published a research post on November 6, 2026, describing components of Claude agent memory, including note-writing, between-session review, linked memory structures and selective retrieval.
How is agent memory different from a context window?
A context window holds information available during a current session. Agent memory is saved information that the agent can read in a later session.
Did Anthropic’s tests prove that smaller memory systems are always better?
No. The post reports particular evaluation results, including a Pokémon comparison with one run for each model. The examples suggest that selective, organized notes can help in some settings, but they do not prove a general rule for all tasks.
What problems did the accompanying account report?
The author described an overfilled index, outdated notes, messages crossing between tasks, loss of detail during handoffs and the need to record rules with dates and reasons. These are practitioner observations, not controlled findings.
Source: ThorstenMeyerAI.com
