Claude Cowork Shares Chat Memory—but Sensitive Topics Stay Opt-In

Anthropic began using the same account-level memory across Claude Chat and cloud-based Claude Cowork tasks on August 25, 2026. Context retained in either mode can now become available in the other, while sensitive subjects remain excluded unless the user separately opts in.

The change is live, not a future preview. Anthropic’s August 25 announcement confirms the two-way connection, topic-level controls and sensitive-memory default; it also limits the Cowork side of the feature to tasks running in the cloud.

What crosses between Chat and Cowork

The shared layer consists of saved memory topics, not a wholesale copy of every previous conversation. Entries can cover a user’s role, current projects, deadlines, working preferences, recurring collaborators, technical conventions and other context Claude considers useful across sessions.

The exchange works in both directions. Context retained during an ordinary chat can inform a later Cowork task, and relevant information encountered during that cloud task can return to Chat. TechCrunch’s launch report also confirms that Claude now adds topics while a conversation is underway instead of waiting to summarize it after the conversation ends.

Anthropic illustrates the change with tasks such as preparing a manager update using known writing preferences or turning conference planning from Chat into a Cowork logistics document without re-entering the headcount, location and speakers. Those are product examples rather than a promise that Claude will retain every detail: the system selects and updates individual memory entries.

Memory remains distinct from past-chat search. Search retrieves relevant conversation material when invoked, whereas memory supplies retained topic entries across eligible sessions. Projects also keep separate memory spaces, so context saved inside one project does not automatically become the memory of another project or a standalone conversation.

Users can inspect, change or erase saved topics

Under Settings > Memory, the Topics area lists the entries Claude has retained. Users can open an entry, correct outdated or inaccurate information, delete it, or tell Claude within a chat what to remember, change or forget. An edit applies to subsequent conversations and eligible cloud Cowork tasks.

Claude’s current memory documentation says the feature is on by default for Free, Pro and Max accounts on the web, Claude Desktop and updated mobile apps; Team and Enterprise organizations must first have it enabled by an owner. The same documentation specifies that cloud Cowork tasks share this memory, while Cowork sessions running locally on a computer do not use it.

Pausing memory preserves existing entries but prevents Claude from using them or creating new ones. Resetting memory permanently deletes all entries, including project memories, and restarts the feature with an empty store. Deleting the conversation that produced an entry is not enough: a related memory can remain until it is separately deleted from Topics or removed through a reset.

Memory entries are also included in account data exports. For organizational accounts, retention and export policies still govern the underlying data even though owners cannot open or edit an individual member’s saved topics.

Sensitive memory has its own consent gate

Claude does not save topics involving health, race, ethnicity, religious beliefs, politics or gender identity by default. A user who wants eligible information in those categories retained must separately enable Include sensitive topics in memory.

The permission applies only from that point forward; enabling it does not scan earlier conversations and add sensitive details retroactively. When Claude saves an eligible sensitive entry, the interface displays a notice so the user can review the memory or revisit the setting. If the option is later switched off, previously saved sensitive entries are removed and new ones are no longer created.

Some categories remain outside memory even after consent. Claude does not retain government identification numbers, financial account numbers, criminal history or immigration status, and it is meant to tell the user when a requested memory update falls into one of those exclusions.

Team and Enterprise accounts add an organizational gate. An owner must permit both memory and sensitive-topic memory, but each member must still opt in personally before those categories can be retained. Owners cannot inspect or edit a member’s individual memories. Memory is unavailable to organizations covered by HIPAA, public-sector or custom data-retention agreements.

Incognito and local sessions remain outside the bridge

Incognito chats neither use existing memory nor create entries for later sessions, and they do not appear in the user’s visible chat history. That is a session boundary rather than immediate deletion: incognito conversations are retained for 30 days by default, or longer under an Enterprise organization’s custom retention policy. Team and Enterprise owners can receive them through organizational exports.

The Cowork boundary is based on where the task runs, not simply which device launches it. Although Cowork spans web and mobile as well as desktop, only cloud tasks participate in the shared account memory described in this update; locally running sessions remain outside it.

The result is one editable memory layer across regular Chat and cloud Cowork, but not one unrestricted archive. Project spaces remain separate, local Cowork and incognito sessions do not join the exchange, and sensitive information still requires an additional decision from both organizational administrators and individual users where those controls apply.

Also read:

• Claude Cowork Is No Longer Mac-Only—Its Agent Now Spans Web and Mobile
• Goldman’s Claude Agents Moved Toward Launch, but Results Stay Undisclosed
• Revolut Is Now a UK Bank, but Some Accounts Are Still Mid-Switch