🔍 Read the full analysis: A Look At AI Agents With Memory: Findings From Anthropic And Our Runs on ThorstenMeyerAI.com
Get smart everyday buys 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, 2026, research post describes how AI agents can save, organize and revise information between sessions. Its examples suggest that structured, selective memory can outperform large collections of notes, but the results are limited evaluations, including single runs. ThorstenMeyerAI.com also reports five operational problems observed in its own Claude-agent fleet.
Anthropic published a research post on November 6 describing how AI agents can write, organize and revise notes that persist between conversations. The post, “The components of agent memory,” presents a framework and results from Anthropic’s own evaluations, while a November 8 report from ThorstenMeyerAI.com describes five practical problems its authors encountered running a small fleet of Claude agents with memory.
Anthropic defines agent memory as knowledge gained through interactions or assigned tasks that the system retains between sessions in a form it can read later. That is different from a model’s context window, which holds information available during the current session. The post also distinguishes memory from changes to model weights, learning within one conversation and summaries used to continue work after a context window fills.
The framework identifies several components: agents decide what is important enough to save, store notes or mark events in session logs, and later revisit those records. Anthropic calls this review-and-revision stage “dreaming.” It can be used to correct inaccurate notes, extract rules or prepare information for future sessions. Memory may also be arranged as linked subject pages, with a small top-level index loaded automatically and other material retrieved when needed.
Anthropic’s examples indicate that writing down more is not necessarily better. In one Pokémon Red comparison at about 10,000 game steps, Claude Sonnet 3.7 created 413 files across 225 folders and had earned one badge. Claude Fable 5 created nine files in two folders, plus a top-level dashboard, and had earned four badges. Anthropic describes one run of each model, so the comparison is an illustration, not a broad performance benchmark.
On a separate database task from Continual Learning Bench 1.0, questions began in fresh sessions and the memory folder was the information carried forward. Anthropic says some models saved untested guesses, while Claude Opus 5 and later models in the comparison recorded rules alongside evidence and warnings about traps. A separate company-document task found that a self-built wiki raised scores from 13% to 48% of graded criteria across 128 sessions. These are Anthropic-reported evaluation results, not independent validation.
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
Persistent memory changes what users can reasonably expect from an assistant. A system that can refer to prior decisions may avoid asking the same questions repeatedly or losing task details when a session ends. But the information it saves can also shape later answers, making accuracy, retrieval and correction central to reliability.
The examples point to a practical trade-off: indiscriminate note-taking can create clutter, while a compact index and task-relevant records may make useful information easier to find. ThorstenMeyerAI.com says its own Claude agents surfaced five recurring issues: indexes fill up, old notes become stale, messages can cross between records, handoffs lose detail, and rules work better when recorded with a date and reason. Those observations are the site’s reported experience, not Anthropic’s evaluation findings.
For organizations adopting agents, the distinction matters because saved notes can outlive the conversation that produced them. Teams will need to know what an agent retains, how it checks records and how users can correct or remove mistakes. The source material does not establish a general reliability advantage or a standard approach that works across systems.
Products Worth Considering
Apple iMac (2024) 24-inch All-in-One Desktop with M4 Chip, 16GB Unified Memory, 256GB SSD, Green
Apple iMac (24-inch, 2024) with M4 Chip, 16GB Unified Memory, 256GB SSD, Silver
Apple MacBook Air 15-inch (2025) with M4 Chip, 24GB Memory, 512GB SSD, Midnight
As an affiliate, we earn on qualifying purchases.
From Session Notes to Memory Systems
Traditional chat assistants generally respond using the current conversation and any information supplied in it. When that session ends, its contents are not necessarily available in a later one. Agent memory adds a separate store that can be read in future sessions, rather than expanding what the model was trained to know.
Anthropic’s post draws an analogy to offline processing of human memories, and connects its “dreaming” concept to earlier work on sleep-time compute proposed by researchers at Letta and UC Berkeley. In Anthropic’s account, this is a design analogy and a way to organize agent processes; it does not establish that AI systems remember in the same way people do.
The source article says Anthropic’s post is a research post, not a peer-reviewed paper, and that several reported results came from single runs. ThorstenMeyerAI.com says its agents also run on Claude and that Anthropic did not review or sponsor its account. Its observations should be read separately from Anthropic’s published evaluations.
““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 Published Evidence
The results described are Anthropic’s own evaluations, and the source says several were single runs. The Pokémon comparison, in particular, does not show whether the difference would persist across repeated trials, other games or different tasks. The company-document result is reported as a change in the share of graded criteria met, but the source does not provide enough information here to assess how broadly it generalizes.
The material also does not establish how often an agent’s stored notes will be wrong in everyday use, how reliably “dreaming” detects errors, or what safeguards are needed when memories contain sensitive or conflicting information. The five lessons from ThorstenMeyerAI.com are attributed to its own experience; the article does not provide independent testing of those observations.
Questions for Wider Testing
The source material does not announce a scheduled follow-up study or product release. The next evidence readers would need is repeated testing across tasks and models, with clear reporting on how memory stores are built, how errors are corrected and how results compare with agents that do not retain notes.
For now, Anthropic’s post offers a framework and company-reported examples rather than a universal recipe. ThorstenMeyerAI.com’s account adds operational cautions from its own Claude-based agents. Whether those approaches produce dependable gains in routine use remains to be established.
Key Questions
What does Anthropic mean by agent memory?
It means information an agent gains through interactions or tasks, saves in a form it can read, and carries into later sessions. Anthropic distinguishes this from the current conversation’s context window and from changes to model weights.
What is “dreaming” in Anthropic’s framework?
Anthropic uses the term for a process between sessions in which an agent reviews logs or transcripts, checks existing notes and may correct or reorganize its memory. The post presents it as a system-design component, not human-like sleep.
Did the Pokémon test prove that fewer memory files are better?
No. The source describes one run each of two models, with different file counts and badge totals. Anthropic uses the example to illustrate selective, structured memory, but it is not enough to establish a general rule across models or tasks.
Who reported the five practical problems with agent memory?
ThorstenMeyerAI.com reported them based on its own small fleet of agents running on Anthropic’s Claude. The site says Anthropic did not review or sponsor its account.
Source: ThorstenMeyerAI.com
