🔍 Read the full analysis: A Field Guide To AI Agent Memory, From 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, 2026, describing how Claude agents can save, organize and correct information across sessions. Its evaluations suggest that compact, structured memory can outperform sprawling notes on selected tasks, but the results are company-reported and include single-run comparisons.
Anthropic published a research post on Nov. 6, 2026, describing how its Claude models can write, organize and revise notes that persist between conversations. The post lays out components of AI agent memory and reports results from company-run evaluations, offering a view into how persistent notes may affect an agent’s performance while leaving open how broadly the findings apply.
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 is different from a model’s context window, which holds information available during a current session, and from changes to model weights made during training. It is also distinct from summaries used to make room in an ongoing conversation.
The research post describes several components: memory notes saved by the agent; salience, or its judgment about what to retain; a between-session review process Anthropic calls “dreaming”; and organized representations, such as linked wiki pages. A memory hierarchy can put a short index or dashboard into every session while leaving other notes available for retrieval when needed.
In one Pokémon Red comparison, Anthropic reported that Claude Sonnet 3.7 created 413 files across 225 folders and had earned one badge at about 10,000 steps. Claude Fable 5 created nine files in two folders plus a top-level dashboard and had earned four badges. Anthropic said each model’s result came from one run. The company also reported that a self-built 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 Structure Affects Agents
Persistent memory could let an agent carry forward user preferences, project details and lessons from earlier work, rather than treating each session as a fresh start. That may help with tasks spread across multiple conversations, but it also creates a new reliability problem: saved information can be incomplete, stale or wrong.
Anthropic’s examples point to a trade-off between storing more and storing useful information. Notes about events visible in the current environment may add clutter, while a short plan, tested rule or warning could support later decisions. The results are relevant to developers building agents and users relying on them for ongoing work, but they do not establish that one memory design will improve every model or task.
The report also describes a review stage called “dreaming,” in which an agent checks stored notes against prior transcripts and revises them. If effective, that process could help correct errors rather than carry them forward. The post’s reported evaluations do not, by themselves, establish how reliably such correction works in routine use.
Popular right now
KitchenAid Artisan 5-Quart Tilt-Head Stand Mixer, Almond Cream
De'Longhi Classic Signature Espresso Machine with Adjustable Milk Frother
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Eucalyptus Green
As an affiliate, we earn on qualifying purchases.
From Chat History to Persistent Notes
For a conventional chat assistant without persistent memory, each new conversation may begin without the previous session’s details. The model can use information in its current context window, but that information is not necessarily available the next day. Agent memory is a separate store that can be loaded or searched later.
Anthropic’s post draws on an analogy with human memory, where some consolidation is thought to happen during sleep or rest. It also builds on the idea of “sleep-time compute,” previously proposed by researchers at Letta and UC Berkeley, according to the source article. Anthropic applies the analogy to agents that review logs between sessions and turn them into notes or structured references.
The evaluation results described here come from Anthropic’s own research post, not a peer-reviewed paper. ThorstenMeyerAI.com says it also operates a small fleet of Claude-based agents and reports practical problems including full indexes, outdated notes, crossed messages and loss of detail during handoffs. The site says Anthropic did not review or sponsor its article.
“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 Results
The post’s evaluations are company-reported, and the Pokémon comparison involved one run per model. The source material does not provide enough detail to determine whether the difference in badges would persist across repeated runs, other games or different tasks. The reported wiki result is tied to a particular set of 25 questions and 128 sessions; its broader performance implications are not established here.
It is also unclear how Anthropic’s memory methods perform in everyday deployments, how often stored notes are corrected successfully, or how the systems handle conflicting information and privacy-sensitive details. The available material does not describe independent replication or a peer-reviewed assessment. The claims should be read as results from Anthropic’s evaluations, not as settled evidence about agent memory generally.
Testing Memory Beyond Company Evaluations
The next step for judging the approach is to see whether the reported gains hold across repeated trials and a wider range of tasks, models and memory designs. Independent evaluations could help establish whether compact dashboards, linked notes and transcript-based correction consistently improve performance.
For users and developers, the near-term issue is how to inspect and update what an agent retains. Anthropic’s post describes mechanisms for organizing and revising memory, but the source material does not set out a general product rollout or a schedule for further results. Until more evidence is available, persistent notes should be treated as useful working records that may need checking—not as a guaranteed or error-free account of prior sessions.
Key Questions
What does Anthropic mean by AI agent memory?
It means information a model gains from interactions or tasks and stores for use in later sessions. It is distinct from the current conversation’s context window.
What is “dreaming” in the research post?
Anthropic uses the term for an agent’s between-session review of logs and transcripts. The agent can use that review to correct or reorganize saved notes and prepare information for future sessions.
Did Anthropic show that fewer memory files always work better?
No. The Pokémon example compares one run of each of two models in a particular game setting. It suggests that a compact, organized memory performed better in that comparison, but does not establish a universal rule.
Are the findings peer-reviewed?
The source describes Anthropic’s results as coming from a research post and company-run evaluations, not a peer-reviewed paper. Independent confirmation is not provided in the material.
What risks come with persistent agent memory?
Notes can be outdated, mistaken or poorly organized, and errors may carry into later sessions. The source article also reports practical issues such as full indexes and lost detail during handoffs.
Source: ThorstenMeyerAI.com
