이미지: @claudeai (X) 화면 갈무리
Summary
- Anthropic announced on August 25 that it has merged the memory systems of Claude chat and Claude Cowork into one
- The system has moved from summarizing conversations only after they end to saving topics in real time as a conversation unfolds, and users can view, edit, or delete anything that's stored
- Sensitive information like health, religion, or politics isn't stored by default, but users can turn it on with a toggle if they want
- 발표일
- 2026년 8월 25일, 앤스로픽 공식 X 계정
- 핵심 변경
- 클로드 챗과 클로드 코워크의 메모리 시스템 통합
- 저장 방식 변화
- 대화 종료 후 요약 → 대화 중 실시간 토픽 추가
- 메모리 관리
- 설정 > Memory 화면에서 항목별 열람·수정·삭제 가능
- 민감 정보 기본값
- 건강·인종·종교·정치·성정체성 등은 기본 미저장, 토글로 활성화 가능
- 절대 저장 금지 항목
- 정부 발급 신분증·사회보장번호·범죄 이력·이민 상태
- 적용 플랜
- Free·Pro·Max, 웹·데스크톱·모바일(iOS·안드로이드는 앱 업데이트 필요)
- 부가 기능
- 타사 AI 계정에서 메모리를 가져오는 'Start import' 버튼 제공
What you say in chat, Cowork now picks up
Anthropic updated Claude's memory system on August 25. Until now, the chat interface and Cowork — the agent that handles tasks on your behalf — each kept their own separate memory. Now the two share a single memory. According to a notice Anthropic posted on X, when you hand a task to Cowork, it already knows the project details you discussed in chat, your manager's preferences, and even last quarter's customer information.
Cowork didn't know chat — so you had to explain everything again
Cowork is Claude's agentic feature that creates files and handles scheduling on your behalf. The problem was that Cowork previously had no way to reference anything discussed in chat. According to TechCrunch's report, even if you'd spent days refining an idea in chat, the moment you handed the actual work to Cowork, you had to brief it from scratch all over again. Anthropic gave examples like drafting a status update for a manager or putting together a conference agenda, explaining that Cowork now already knows details like headcount, host city, or speaker lineup, exactly as discussed in chat. With this integration, chat and Cowork no longer function as two separate products under one roof — they behave more like a single, continuous assistant.
How memory builds up, and what it leaves out
The way memory is stored has also changed. TechCrunch reported that instead of summarizing a conversation into memory only after it ends — the old approach — Claude now adds topics to memory in real time as the conversation happens. That means context can flow back and forth between chat and Cowork even while a conversation is still ongoing.
According to screenshots of the settings screen that have surfaced, stored memory is organized into two categories: "You" and "Topics." "You" holds information like the account owner's job title or preferred response style, while "Topics" logs specific subject-based entries with dates attached — things like meal planning, a customer pitch, half-marathon training, or prepping for a quarterly offsite. Hovering over any entry reveals a trash icon that lets you delete it on the spot.
| Category | Default | User control |
|---|---|---|
| Health, race, religion, politics, sexual identity | Not stored | Can be enabled via the "Include sensitive topics" toggle |
| Government ID, Social Security number, criminal record, immigration status | Not stored | Cannot be enabled even via toggle (prohibited under terms of service) |
| Notification when sensitive topics are stored | — | App notifies the user each time this happens |
How to use it
Memory settings can be found in the Claude app's Settings. Click Memory under the Customize section in the left sidebar to open the screen.
- Go to Settings > Customize > Memory to find two toggles: "Search and reference chats" and "Generate memory from chats." Both are on by default, letting Claude search past conversations and generate new memory entries.
- Below that, turning on "Include sensitive topics in memory" allows Claude to store sensitive content such as health conditions or religious beliefs in memory.
- Further down, you'll see lists split into "You" and "Topics," where you can check when each entry was last updated and delete anything you no longer want kept.
- If you want to bring over context from another AI service, click "Start import" under "Import memory from other AI providers." Anthropic provides a prompt you can use to import memory from your existing account elsewhere.
According to TechCrunch, the feature is turned on by default across Free, Pro, and Max plans on web, desktop, and mobile, though iOS and Android users need to update to the latest app version for it to take effect. Practical use cases include organizing recurring weekly meeting agendas, drafting a new pitch deck using last quarter's customer information, or having Cowork instantly turn a meal plan discussed in chat into a grocery list.
Anthropic had previously rolled out updates to memory, domain controls, and a session viewer for Claude's managed agents. This latest announcement looks like an extension of that same push into consumer-facing products like chat and Cowork.
Editor's view
The real story here isn't where memory is stored, but when it's captured. Under the old structure, where a summary only formed after a conversation ended, other products had no way to tap into context from a conversation still in progress. Switching to real-time topic capture is effectively a decision to treat chat and Cowork as a single process, and that's a different philosophy from OpenAI or Google, which still run their chatbot and agent products as separate lines.
Anyone who's put an agentic product of similar scale to work has run into the same wall repeatedly: spending hours refining requirements in chat, only to have to retype everything from scratch the moment you hand the task off. Removing that repeated step matters more than it might sound, especially for teams working on a single project over several weeks — the time saved there will be noticeable.
The point worth watching in practice is the sensitive-info toggle. Having it off by default is a safeguard, but for organizations where multiple people share the same Claude account, one person flipping that toggle changes the memory policy for the entire team. For companies here adopting this feature, simply having an account admin periodically check that toggle setting could meaningfully cut down the risk.
In the coming weeks, it wouldn't be surprising to see OpenAI or Google make similar announcements breaking down the memory divide between their chatbots and agents. It looks like memory integration is already becoming the next standard feature in the race among agentic products.
Comments