TL;DR
Anthropic published a research post on November 6, 2026, describing how Claude agents can save, organize and correct information between sessions. Its reported evaluations suggest that compact, structured notes can outperform large collections of session details, though the results are limited and not peer-reviewed. The article’s authors also describe operational problems from their own agents, including stale notes and lost handoff details.
Anthropic published a research post on November 6, 2026, outlining how its Claude models can save, organize and revise information between conversations. The post describes components of agent memory and reports evaluations in which more selective, structured notes performed better than a much larger collection of files, a development relevant to users who expect agents to carry instructions and task knowledge from one session to the next.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Anthropic defines agent memory as knowledge a model gains from interactions or assigned tasks and retains between sessions in a form it can read later. Its account distinguishes this from a model’s context window, which contains information available during the current session, and from model training changes. It also separates persistent memory from summaries that replace earlier conversation turns to make room in an active session.
The post describes several parts of a memory system: notes selected for saving; a salience judgment about what is likely to matter again; a between-session review process Anthropic calls “dreaming”; navigable representations such as linked wiki pages; and a hierarchy in which a short overview can load automatically while other notes are searched as needed. Anthropic says the review process can compare stored notes with session records and revise or add information.
In one reported Pokémon Red evaluation, Anthropic compared single runs of Claude Sonnet 3.7 and Claude Fable 5 at about 10,000 game steps. The older model had created 413 files in 225 folders and earned one badge; the newer model had created nine files in two folders plus a top-level dashboard file, and earned four badges. Anthropic argues that repeated on-screen details need not be stored, while strategies and tested rules may be useful to preserve. The comparison is a limited example, not a broad benchmark.
What agents keep between sessions
Anthropic’s research post describes memory as selected, retrievable notes that can be reviewed and revised. Its examples are promising, but limited.
Memory is more than chat history
A context window serves the current session; persistent memory stores selected information for later retrieval.
- Context window holds available conversation and supplied material.
- Session summaries make room as the conversation grows.
- Saved notes can carry useful knowledge forward.
- Indexes and linked pages help retrieve relevant details.
A selective memory system
Anthropic describes choosing what to save, reviewing it later, and retrieving notes as needed.
Save information likely to matter again, such as strategies or tested rules.
“Dreaming” checks stored notes against session records and may revise or add them.
Load a short overview or search other notes for task details.
A striking, narrow comparison
In one Pokémon Red evaluation at about 10,000 game steps, the newer model used fewer files and earned more badges.
| Model | Memory files | Badges |
|---|---|---|
| Claude Sonnet 3.7 | 413 files in 225 folders | 1 |
| Claude Fable 5 | 9 files in 2 folders, plus a dashboard | 4 |
Reliability still needs work
More stored material can bury useful instructions, while stale or mistaken notes can mislead later sessions.
Check freshness
Changed deadlines and preferences need reliable updates so agents do not follow obsolete notes.
Treat results cautiously
The examples are company-reported, limited evaluations, not proof that compact memory works better everywhere.
Memory Design Shapes Agent Reliability
Persistent memory can affect whether an agent follows a changed deadline, recalls a user preference or repeats work after a session ends. The post frames memory as an active system that selects and retrieves information, rather than a transcript archive. That distinction matters because more saved material can make useful instructions harder to find and can preserve details that no longer apply.
The post also highlights a reliability problem: stored information can be mistaken or outdated. A review step that checks notes against prior records offers one possible correction mechanism, but it does not by itself establish that an agent will identify every error. The authors’ separate account of their own agents describes crowded indexes, stale notes, messages crossing and handoffs losing detail. Those observations are experience reports, not controlled findings.
Products Worth Considering
Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite
Amazon Kindle Scribe Colorsoft 64GB with 11-Inch Color Display and Premium Pen, Graphite
EcoFlow DELTA 3 Max Portable Power Station, 2048Wh LFP Battery
As an affiliate, we earn on qualifying purchases.
From Chat History to Persistent Notes
Many chat assistants historically began each new conversation without access to the prior session. A context window lets a model work with the current conversation and supplied material, but information outside that window is unavailable unless it is stored and brought back later. Agent memory addresses that gap by keeping selected information in files or another retrievable store.
Anthropic’s post draws an analogy to offline memory processing in neuroscience and refers to “sleep-time compute,” an idea previously proposed by researchers at Letta and UC Berkeley. The analogy is a design inspiration, not evidence that AI systems remember in the same way people do. Anthropic’s post is a company research account, and the source article says the company did not review or sponsor its own experiments.
The article also reports a separate evaluation involving a self-built wiki for questions about company documents: across 128 sessions, the wiki raised the share of graded criteria met from 13% to 48%. These figures are attributed to Anthropic’s post. The provided material does not give enough information about the full evaluation setup, repeated trials or uncertainty to treat the result as a general performance estimate.
““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 Reported Evaluations
Anthropic’s findings are presented in a research post, not a peer-reviewed paper, and the source says several results are based on single runs. The Pokémon comparison involves two model versions in one game setting; it does not establish that smaller memory stores will perform better across other tasks or models. The supplied account also does not provide enough methodological detail to assess how broadly the company-document wiki result would transfer.
It remains unclear how consistently an agent can distinguish durable facts from temporary instructions, detect contradictions, or prevent an outdated note from influencing a later answer. The source does not report a general error rate for the memory correction process or establish how much these techniques improve real-world reliability over repeated use.
Testing Memory Across More Tasks
The next useful evidence would include repeated evaluations across different tasks, model versions and memory setups, with clear reporting of errors, retrieval success and the cost of maintaining notes. Tests of correction processes would also need to show whether agents catch stale or conflicting information before it affects their work.
For now, Anthropic’s post offers a description of its memory design and a set of company-reported examples. The publisher says its own agents continue to surface practical issues with indexing, freshness and handoffs, but the source provides no timeline for further experiments or product changes.
Key Questions
What does Anthropic mean by agent memory?
It means information a model gains from interactions or tasks, saves in a form it can read later, and carries between sessions.
How is agent memory different from a context window?
A context window holds information available in the current session. Agent memory stores selected information for retrieval in a later session.
Did Anthropic show that smaller memory stores always work better?
No. The post reports a single Pokémon game comparison in which Claude Fable 5 used fewer files and earned more badges than Claude Sonnet 3.7. That example does not establish a general rule across tasks.
What does Anthropic mean by “dreaming”?
It is the post’s name for a between-session review in which an agent checks its memory against session records and may correct or add notes.
Has the memory approach been independently validated?
The source describes Anthropic’s work as a research post rather than a peer-reviewed paper. The provided material does not establish independent validation of the reported results.
Source: Thorsten Meyer AI
