Until now, Claude had two heads. One remembered what you discussed in chat. The other lived in Cowork and started every task from scratch. As of August 25, that’s over: chat and Cowork share the same memory.
One memory instead of two
The practical upshot: you explain something once, and both sides know it. Talk through the agenda for a conference in chat, and when Cowork builds the budget doc afterward, it already knows the headcount, the city, and the speakers. Explain once how your team defines its metrics, and every quarterly deck Cowork builds after that uses them. No rebriefing.
Anthropic also changed when Claude remembers things. It used to summarize a conversation after it ended. Now Claude adds topics to memory while you chat. Say a deadline slipped to September, and your next conversation already knows - without you saying “remember this.”
You see and edit everything
The part I like most: everything Claude has saved about you sits as a list of short files under Topics in your Memory settings. You can read, edit, or delete each one. If your old company name is wrong somewhere, fix it in one place - and from then on it’s right in every conversation.
That’s more honest than the black-box version, where you never quite know what the model has piled up about you.
Sensitive topics stay your call
By default, Claude does not store sensitive subjects: health, ethnicity, religious or political beliefs, gender identity. But you can turn that on deliberately. Then Claude will remember, say, your gluten allergy when it suggests recipes. Each time it saves something on one of these topics, you get a notice, and nothing gets added retroactively.
A few things stay off-limits even with the setting on: Social Security numbers, government ID numbers, criminal history, immigration status. There, Claude simply tells you it can’t store that.
Availability
On Free, Pro, and Max, memory is on by default across web, desktop, and mobile - except for storing sensitive topics, which you have to switch on yourself. On Team and Enterprise, admins decide whether the feature is even available, and it’s off for individual users to start.
My take
Merging memory across tools is the logical step once Cowork is meant to run tasks in the cloud. An assistant that forgets half the backstory the moment it switches from talking to doing is half an assistant. What I find almost more interesting is the editable files. Being able to read and correct your own profile means handing over less control - and in the end, that’s what memory is really about: trust.
Sources: