One memory now spans Claude chat and Cowork

Anthropic has merged Claude's two separate memories, one in chat and one in Cowork, into a single store that spans both surfaces. The update went live on Tuesday, according to Thenewstack, and it also changes the moment at which Claude writes to memory.

At a glance

• The two stores were previously disconnected: Cowork's memory sat inside individual projects, so a task there could not draw on what Claude had learned about a user across months of chat.
• Memory is updated during a session rather than only once a chat has ended; Cowork sessions, unlike chat exchanges, can run for hours or several days.
• Some categories never enter memory even with the sensitive-topics setting enabled: Social Security and government ID numbers, criminal history, immigration status, and anything that breaches Anthropic's Acceptable Use Policy.

The merge reads as another step in Anthropic's gradual collapsing of the line between chat and Cowork, two surfaces that until now kept separate state. The practical consequence is that context gathered in casual conversation starts shaping agentic work, which raises the cost of a wrong entry, since memory files can be deleted but not edited by hand. The split defaults, on for consumer plans and off for Enterprise and Teams, likely reflect where the compliance questions land.

Chat context now reaches Cowork tasks, and what surfaces in Cowork returns to chat

Anthropic says Claude will now remember what it knows about a user no matter where the conversation happens, and that what comes up in Cowork carries back to chat. By Anthropic's account, the shared context means users will not have to re-explain themselves as often in Cowork, where deeper work tends to happen.

Anthropic gives two illustrations. In one, a request to draft an update for a manager is answered by a Claude that already knows who that manager is and how she likes updates written. In the second, an agenda brainstormed in chat feeds a conference budget and logistics document built in Cowork, with the headcount, city and speakers already in place.

Claude also updates memory while a task is in progress rather than waiting for a chat to end, which was the previous behaviour. Anthropic says months of accumulated context, such as quarterly priorities and project status, is available the moment a task is handed to Cowork.

Health data, religious beliefs and gender identity stay out of memory by default

By default Claude does not store what Anthropic classifies as sensitive, including health data and information about religious beliefs, race, ethnicity and gender identity. A setting turns those categories back on for users who want them remembered, but the control is all or nothing: there is no way to admit some sensitive topics and exclude others.

Other categories are never stored at all: sensitive identification numbers, including Social Security and government ID numbers, criminal history, immigration status, and anything that violates Anthropic's Acceptable Use Policy. Memory is stored as small files organized by topic, which users can delete but not edit directly.

Memory changes are made by chatting with Claude from the settings menu

Direct edits go through Claude itself: from the settings menu, a user can chat with the model to change what is stored. The unified memory is on by default for Free, Pro and Max plans across web, desktop and mobile, with sensitive information excluded unless the setting is changed; Enterprise and Teams have all memory features off by default.

An Anthropic spokesperson notes that Claude remains ad-free, so nothing in memory is used for ad targeting. OpenAI, meanwhile, recently launched Computer History, an optional feature that logs what a user does on a Mac, including inside Claude, and makes it recallable.

What the settings do not cover

The sensitive-topics control remains a single switch, with no option to admit some categories and block others, and Anthropic's description does not say whether finer-grained control is planned. Nor does it set out how the unified memory will be handed to Enterprise and Teams administrators beyond the off-by-default state. Deletion stays the only direct action a user can take on a memory file.