Claude Memory Merge: Chat, Cowork, Defaults and Sensitive Data

One memory instead of two

Claude chat and Cowork had kept separate memory stores since Cowork launched in January. Anthropic's post describes the fix as combining them: whatever Claude learns in one now shows up in the other, so a manager's name mentioned in chat is already known when Cowork drafts a status update, and a preference set while delegating a Cowork task carries back into an ordinary conversation. [1]

Anthropic also changed how a memory gets written. Earlier versions summarized a conversation once it ended. Memory now updates while a chat is still running, adding topics as they come up rather than waiting for a close. [1][2]

The unified memory is on by default for Free, Pro and Max plans across web, desktop and mobile, though Cowork itself is not offered on the Free tier. Team and Enterprise customers get it only when an administrator turns it on, off by default there. Claude Code keeps its own separate memory and was not folded into this merge. [1][3]

What stays out unless a user asks for it

By default Claude excludes health, race and ethnicity, religious belief, political views and gender identity from memory, even though the rest of a conversation is otherwise fair game for it. A user can opt a category back in, Anthropic's example being someone who wants dietary restrictions remembered for meal planning, and gets a notice each time something in a sensitive category is saved afterward. [1]

A separate, non-optional list never enters memory at all: government identification numbers such as a Social Security number, criminal history, immigration status, or anything that would violate Anthropic's acceptable use policy. Turning on sensitive topics widens what can be remembered; it does not reach this second list. [1]

Users can read, edit or delete any saved topic, pause memory, or reset it outright through a settings page, and the change is prospective: opting into a sensitive category does not retroactively pull it from earlier conversations. The Register reported that Anthropic confirmed there is no way to keep chat and Cowork memory separate short of that reset, a different account, disabling memory, or incognito chats. [1][3]

Why the seam between surfaces is the thing to watch

Atlas interpretation: Anthropic now ships Claude through several distinct surfaces: chat, Cowork, Claude Code, and Claude in other companies' products through partnerships and APIs. Each surface previously accumulated its own picture of a user, which meant repeating context every time a task moved from one to another. Merging chat and Cowork memory is a bet that a single assistant identity, portable across modes, is worth more than the isolation those separate stores provided by default. [1][2]

Atlas interpretation: That bet sharpens a question that gets harder as an assistant's surfaces multiply rather than easier: how much does one account's Claude know about a person, and who else can see it. The sensitive-topic exclusions and the notice-on-save mechanic are Anthropic's answer, but the underlying tradeoff, one memory that follows a user everywhere against several memories that stay local to where they were made, is the same one every assistant vendor with more than one product surface will eventually have to make. [1][3]