Claude memory crosses into Cowork — off by default for teams

Anthropic shipped Claude memory into Cowork today, and the interesting part is not the feature. It's the default. Memory is on out of the box for Free, Pro, and Max plans, and off by default for Team and Enterprise organizations, where only an Owner or Primary Owner can turn it on. Which means the personal account someone signed up with over lunch is already accumulating context about your business, and the paid org account your policy actually covers is not.

What actually happened

Per Anthropic's release notes dated August 25, the update is "Memory in Claude Cowork, editable topics, and a sensitive topics setting." What Claude picks up in a chat is now available when you hand it a task in Cowork running in the cloud, and what surfaces during that Cowork task carries back into chat. TechCrunch reports the system now writes topics to memory during a conversation rather than summarizing at the end, which is why the two surfaces stay in step.

The boundaries, from Anthropic's own documentation:

• Local Cowork sessions do not use memory. Cloud sessions do. Where the agent runs determines what it remembers.
• Sensitive topics are excluded by default — health, race, ethnicity, religious belief, politics, gender identity — with a new toggle to opt in.
• Some things are never stored at all: government ID numbers, financial account numbers, criminal history, immigration status. Claude tells you when it refuses to save something.
• Users read, edit, and delete their own memory under Settings > Memory. Admins cannot read individual users' memories — but memory data does land in organizational exports.

Why shared agent memory matters for your business

Your handoff just became a data flow. Research in chat, execution in Cowork used to be two separate contexts with a copy-paste between them. That copy-paste was a control point, and it's gone. Whatever your team pastes into a chat window — a client's margin structure, an unsigned quote, a vendor's rate card — is now standing context for an agent that takes actions.

"Off by default" is not the same as "handled." Team and Enterprise owners get to decide, which is the right design. It also means the decision is now sitting in somebody's queue unmade, while the shadow-IT personal accounts already defaulted to yes. Find out which accounts your people actually use before you write the policy.

Admins can't read it, but exports can. That asymmetry is worth understanding before an employee leaves, before a discovery request, before a client asks what your tooling retains about their account. Memory you cannot inspect day to day but can produce in bulk is a records-retention question, not a feature question.

Sources: Anthropic release notes, Anthropic Help Center, TechCrunch.