🔍 Read the full analysis: How AI Agent Memory Holds Up In Practice: What We Learned 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’s November 6 research post describes how Claude agents can save notes between sessions, correct them and organize them for later use. Its company-run tests suggest that more memory is not always better, while a maintained wiki helped on one document-question benchmark. The results are not peer-reviewed, and some comparisons were single runs.
Anthropic published a research post on November 6, 2026, describing how Claude agents can retain and organize information between sessions—a capability that can shape what an agent remembers when a new task begins. The post reports company-run tests in which a smaller, more structured set of notes performed better than a much larger collection in one game comparison, while a wiki improved results on a separate document-question task.
Anthropic defines agent memory as knowledge gained through interactions or assigned tasks that a model retains between sessions in a form it can read later. That is distinct from the model’s context window, which holds information available during a single session, and from changes to model weights during training. The post also distinguishes memory from summaries that replace earlier conversation turns within a continuing task.
The system Anthropic describes has several parts: memory notes written during or after a task; a salience judgment about what is worth keeping; an offline review process it calls “dreaming,” which checks notes against transcripts; and structures such as linked wiki pages. A hierarchy can keep a small index available at the start of a session while leaving detailed notes to be searched only when relevant.
In one Pokémon Red test, Anthropic compared a single run each of Claude Sonnet 3.7 and Claude Fable 5 at about 10,000 steps. Sonnet 3.7 had created 413 files across 225 folders and earned one badge; Fable 5 had nine files in two folders plus a top-level dashboard and had earned four badges. Anthropic says the more selective notes emphasized goals, plans and tested rules rather than repeatedly recording information visible in the game. The company also reports that a wiki raised performance on 25 questions about company documents 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
Why Memory Design Changes Agent Work
For users, the practical issue is not simply whether an agent can save information, but whether it can retrieve the right information and distinguish a reliable instruction from an outdated guess. A note that a deadline is Friday could help a later session; an old note or an unverified assumption could instead send work in the wrong direction. Memory quality affects continuity, but retaining more material does not by itself guarantee more accurate decisions.
The results also point to a design trade-off for teams deploying agents. A large, unstructured archive can consume attention and make relevant facts harder to find. A concise index, linked topic pages and records of how a rule was tested may make retrieval more useful. These are possible lessons from Anthropic’s reported evaluations, not proof that one memory architecture will work best across agents or real-world tasks.
The source article adds operational observations from its own small fleet of Claude-based agents: indexes can fill up, memories can become stale, messages can cross, and handoffs can lose detail. The publisher says Anthropic did not review or sponsor those observations. They are practical reports, not independently verified evaluation results.
Products Worth Considering
Apple iPad Pro 12.9-inch (2017), 512GB Wi-Fi, Space Gray
Samsung 49-Inch Odyssey G9 G91F Curved Dual QHD Gaming Monitor, 144Hz
Anker SOLIX C1000 Gen 2 Portable Power Station, 1,024Wh
As an affiliate, we earn on qualifying purchases.
From Session Context to Saved Notes
Many chat assistants historically treated each conversation as a fresh session, with no saved notes carried forward. An agent memory system changes that arrangement by writing selected information to storage and loading or searching it later. The model’s context window remains the information available on its “desk” during a particular session; saved memory acts more like a filing system that can be consulted in a later one.
Anthropic’s post draws an analogy with human memory, where some sorting and consolidation is thought to occur during rest. It connects its “dreaming” process to the idea of sleep-time compute previously proposed by researchers at Letta and the University of California, Berkeley. The post’s examples describe a system that revisits interaction logs between sessions to revise notes, rather than treating every saved statement as permanently reliable.
The source article says Anthropic’s reported evaluations are from a research post, not a peer-reviewed paper, and that several results are single runs. Its evidence should therefore be read as an account of early tests and design approaches, rather than a broad comparison of 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, in its definition of agent memory
Limits of the Reported Tests
The figures come from Anthropic’s own evaluations, and the source describes some comparisons as single runs. The Pokémon result compares different model versions in one run each, so it does not establish that memory organization alone caused the difference in badges. The reported wiki improvement is limited to 25 questions about company documents and a 128-session evaluation; the source does not provide enough detail here to judge how well it transfers to other tasks.
It is also unclear how these approaches perform across a wider range of models, users and long-running deployments, or how reliably an agent can detect a false or outdated note. The source article reports issues seen in its own fleet but does not give counts, rates or a controlled comparison for them. No broad independent validation of the post’s findings is cited.
What Further Testing Needs to Show
The source material does not announce a follow-up release or a scheduled milestone. The next evidence readers should look for is replication across more runs, clearer comparisons that separate model capability from memory design, and tests on tasks beyond games and a fixed set of document questions.
For people building or using agents, the reported examples suggest watching whether notes are checked against their source, whether outdated instructions can be corrected, and whether the agent can find relevant details without loading an entire archive. Anthropic’s post outlines a possible architecture; evidence about its reliability in broader, ongoing use remains to come.
Key Questions
What did Anthropic publish?
Anthropic published a research post on November 6, 2026, called “The components of agent memory.” It describes how Claude agents can save, review and organize information for use in later sessions.
How is agent memory different from a context window?
A context window holds information available during a session. Agent memory is information saved between sessions and read or searched later. Anthropic also distinguishes both from model training and from summaries used to continue work within a task.
Did the tests show that less memory is always better?
No. One Pokémon comparison favored a smaller, more structured set of notes, but it involved one run per model. The results do not show that less memory is always better across tasks or systems.
How strong is the evidence?
The reported results come from Anthropic’s evaluations in a research post, not a peer-reviewed paper. The source says several results are single runs, so independent replication and broader testing would be needed to establish how broadly the findings apply.
Source: ThorstenMeyerAI.com
