🔍 Read the full analysis: Running AI Agents With Memory: What Worked And What We Learned on ThorstenMeyerAI.com
Get audio and creator gear 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 November 6 describing how Claude agents can save, organize and correct information between sessions. Its examples suggest that selective notes and searchable structure can help on some tasks, while an independent operator’s experience highlights risks including stale or misdirected memories.
Anthropic published a research post on November 6, 2026, describing how its Claude models can save and use notes across sessions, and reporting results from internal memory tests. The post offers a look at a developing approach to persistent AI agents: keeping information between conversations while deciding what to store, retrieve and correct. The findings are from Anthropic’s own evaluations, not a peer-reviewed paper, and some comparisons involved a single run.
Anthropic defines agent memory as knowledge gained through interactions or assigned tasks and retained between sessions in a form the model can read later. That is different from a model’s context window, which holds material available during a current conversation, and from model weights changed through training. The post also distinguishes persistent memory from summaries that replace earlier turns within an ongoing task.
The post describes several components: notes written during or after a task; a process for judging which details are worth keeping; a between-session review Anthropic calls “dreaming”; and structured representations, such as linked wiki pages. It also describes a memory hierarchy in which a short top-level index is available at the start, while other material is searched or loaded only as needed.
In one Pokémon Red comparison, Anthropic said Claude Sonnet 3.7 wrote 413 files across 225 folders and had earned one badge at roughly 10,000 game steps. Claude Fable 5 wrote nine files in two folders and a top-level dashboard, and had earned four badges. Anthropic reported one run for each model, so the example is illustrative rather than a broad performance comparison. The post’s separate database-task example describes models that differed in whether their saved notes included tested rules and warnings, rather than unverified guesses.
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 Design Shapes Agent Reliability
Persistent memory can make an agent more useful when a task spans multiple sessions: users may not have to repeat preferences, project details or work already completed. But saved information also becomes part of the agent’s future decision-making. Wrong, stale or poorly organized notes can steer later work off course, while excessive notes may make relevant details harder to find.
Anthropic’s examples point to a design trade-off, not a guarantee that memory improves every agent. The Pokémon comparison suggests that more files did not correspond to better progress in that run. For users and developers, the practical question is whether a memory system preserves useful, verified information and makes it easy to retrieve—not simply whether it stores more.
Products Worth Considering
Apple MacBook Air 15-inch (2025) with M4 Chip, 24GB Memory, 512GB SSD, Midnight
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 24GB Memory, 1TB 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
As an affiliate, we earn on qualifying purchases.
From Blank Sessions to Saved Notes
Many chat assistants have historically treated each conversation as a separate session, with no automatic carryover of details from earlier chats. Agent systems that save files or other records between sessions change that arrangement: the model can consult its own accumulated notes when a new task begins.
Anthropic’s post draws an analogy to human memory, where some sorting and consolidation happens away from the immediate experience. It connects this idea to “sleep-time compute,” previously proposed by researchers at Letta and UC Berkeley. Anthropic presents the analogy as a way to think about agent design, not evidence that machine memory works like human memory.
A separate November 8 account from Thorsten Meyer describes operating a small fleet of Claude-based agents. Meyer says the experience exposed five practical problems: indexes can fill up, old notes can mislead, messages can be confused, handoffs can lose detail, and operating rules are more useful when saved with a date and reason. The account is an operator’s experience, not an independent controlled evaluation; it says Anthropic did not review or sponsor it.
“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 Published Evidence
The results described are Anthropic’s own evaluations, published as a research post rather than a peer-reviewed paper. The Pokémon example compares one run of each named model, and the source material does not establish whether the result would hold across repeated runs, other games or different tasks. The database example is described, but the supplied material does not include its full scores or all experimental details.
It is also unclear how consistently the memory techniques perform outside the tested settings, how much they add in time or computing cost, and what safeguards prevent errors from being preserved. Meyer’s five lessons are based on his team’s use of agents, with no sample size or controlled baseline provided. They offer practical observations, not a measured estimate of how often these failures occur.
Testing Memory Beyond Examples
Anthropic’s post lays out components and selected evaluations, but the material available here does not announce a release date, product change or next research milestone. Further evidence would need to show how memory methods perform across repeated trials and varied tasks, including whether agents can reliably detect outdated or incorrect notes.
For developers using persistent agents, the immediate work is operational: inspect what gets saved, retain sources or reasons for key rules, and check whether later sessions retrieve the right information. The next question for readers is not simply whether an agent remembers, but whether its memory can be verified, corrected and kept relevant as tasks and circumstances change.
Key Questions
What did Anthropic announce?
Anthropic published a research post on November 6, 2026, describing components for Claude agent memory, including saved notes, memory review, structured representations and layered retrieval. The source does not describe a product launch.
How is agent memory different from a context window?
A context window holds information available during a current session. Agent memory is information saved in a form the model can read in later sessions, according to Anthropic’s definition.
Did the tests prove that fewer memory files are better?
No. Anthropic reported a Pokémon comparison in which one run of Fable 5 used fewer files and had earned more badges than one run of Sonnet 3.7. That example does not establish a general rule across models or tasks.
What risks come with persistent agent memory?
Saved information may become stale, incorrect or difficult to retrieve. Thorsten Meyer also reported problems with full indexes, confused messages and detail lost during handoffs in his own operation of Claude-based agents.
Was Anthropic’s research independently verified?
The supplied material identifies the results as Anthropic’s own evaluations in a research post, not a peer-reviewed paper. It does not provide independent replication of the reported examples.
Source: ThorstenMeyerAI.com
