🔍 Read the full analysis: AI Agents And Memory: Connecting Anthropic’s Research To Practice 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 published a research post on Nov. 6 describing how AI agents can save, organize and revise information between sessions. Its company-run tests suggest that selective, structured memory can help on some tasks, but the results include single runs and are not peer-reviewed.
Anthropic published a research post on Nov. 6 explaining how Claude agents can store information between conversations, correct their notes and retrieve them later. The post describes memory as a separate system from a model’s active conversation context and reports company-run tests in which organized notes helped on selected tasks; the findings are not a peer-reviewed paper, and some evaluations were single runs.
Anthropic’s post, titled The components of agent memory, defines agent memory as knowledge gained through interactions or assigned tasks that is retained between sessions in a form the model can read later. It distinguishes that from a context window, which contains material available during the current session, and from summaries used to continue work within a conversation. Memory also does not mean changing a model’s trained weights.
The proposed system includes memory notes, a process for deciding what matters, and a step Anthropic calls “dreaming.” During this between-session step, an agent can compare notes with transcripts, correct errors, form new representations or decide what to load next time. Anthropic also describes structured representations such as a linked wiki and a hierarchy in which a small set of notes is loaded automatically while other material is searched only when needed.
In a Pokémon Red test, Anthropic compared one run each of Claude Sonnet 3.7 and Claude Fable 5 at about 10,000 game steps. Sonnet 3.7 created 413 files across 225 folders and had earned one badge; Fable 5 created nine files in two folders plus a top-level dashboard and had earned four. The post also reports that a self-built wiki raised performance on 25 questions about company documents from 13% to 48% of graded criteria over 128 sessions. These are results from Anthropic’s own evaluations, not independent benchmarks.
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 Changes What Agents Retain
Persistent memory addresses a practical limitation of conversational AI: information from one session may not be available in the next. For agents assigned work over multiple sessions, saved preferences, procedures and task history could reduce repeated instructions and help maintain a plan. But the system’s value depends on what it saves and retrieves, not simply on how much storage it has.
Anthropic’s examples illustrate that distinction. Its Pokémon comparison suggests that more files did not correspond to better progress in those particular runs; the newer model’s smaller set included a dashboard with its goal and plan. The company-document test suggests that a structured wiki can make relevant information easier to locate. Neither example establishes that the same approach will improve performance across models, tasks or users.
Memory also creates a new source of error. A stale or mistaken note can influence later answers after the original conversation is gone from view. Anthropic’s “dreaming” process is intended to check and revise stored information, but the post’s descriptions do not establish how reliably this works in general. For people and businesses using agents, memory management becomes part of evaluating the system: users may need to know what is retained, how it is corrected and when it is used.
Products Worth Considering
Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, 12MP Center Stage Camera, and Wi-Fi 7, Sky Blue
Apple MacBook Air 15-inch (2025) with M4 Chip, 24GB Memory, 512GB SSD, Midnight
Sony WH-1000XM6 Wireless Noise-Canceling Headphones, Black
As an affiliate, we earn on qualifying purchases.
From Context Windows to Saved Notes
In a standard chat session, a model responds using the material available in its current context window, including the conversation and any supplied documents. When a new session begins, that material may not carry over. An agent memory system adds a store the model can consult later, making continuity possible without keeping every past exchange in the active context.
Anthropic connects its design to the idea that useful memories can be consolidated offline rather than recorded indiscriminately as events happen. Its post references “sleep-time compute,” a concept proposed by researchers at Letta and the University of California, Berkeley. The post applies that general idea to agents: save potentially useful experience, review it between sessions, then prepare a smaller or better-organized set of information for future tasks.
The source article also describes lessons from a small fleet of agents run by its author, including risks from full indexes, outdated notes, crossed messages and lost detail during handoffs. Those observations are the author’s operational experience, not results from Anthropic’s study. The article says Anthropic did not review or sponsor that account.
““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”
How Reliable Are These Results?
The reported outcomes come from Anthropic’s own evaluations, published as a research post rather than a peer-reviewed study. The Pokémon comparison involved one run per model, so the results do not show how consistently either system would perform across repeated trials. The company-document result is reported as a change from 13% to 48% of graded criteria over 128 sessions; the post’s summary does not establish how the method performs on other document collections or tasks.
It is also not clear from the supplied material how often memory notes become inaccurate, how well the correction process catches those errors, or what safeguards govern sensitive information stored between sessions. The post describes components and examples, but does not provide evidence here of a universal memory design or a general performance guarantee.
Testing Memory Across More Tasks
The next question is whether the described methods hold up across repeated and independent evaluations. Useful follow-up evidence would include results across more tasks and model runs, clearer comparisons between memory designs, and measurements of errors caused by outdated or incorrect notes. Independent testing could also clarify how much performance gains depend on a particular model, task setup or memory store.
For teams deploying agents, the immediate practical step is to treat retained notes as information that needs review, rather than assuming that saved memory is accurate. Anthropic’s post offers a framework—record, organize, revisit and retrieve—but the available evidence does not yet settle how well each component works in broader use.
Key Questions
What does Anthropic mean by agent memory?
It means information an agent gains from interactions or tasks, saves between sessions and can read later. Anthropic distinguishes this from the model’s active conversation context and from changes to its trained weights.
What is “dreaming” in Anthropic’s description?
It is Anthropic’s term for reviewing logs and transcripts between sessions to correct notes, form new memory representations or prepare information for later use.
Did Anthropic prove that less memory works better?
No. The post reports one Pokémon Red run for each of two models, with different file counts and badge totals. That comparison is suggestive but does not establish a general rule.
Are the reported findings peer-reviewed?
No. The supplied source describes the results as Anthropic’s own evaluations in a research post, not a peer-reviewed paper. Some results came from single runs.
What risks come with persistent agent memory?
Saved notes can be stale, incorrect or incomplete and may influence later work. The post describes a correction process, but the supplied results do not establish how reliably it prevents such errors.
Source: ThorstenMeyerAI.com
