Anthropic is connecting Claude’s conversational and task-oriented experiences through a shared memory system. Information saved during ordinary Claude chats can now be available in Cowork, and details learned while Cowork completes a project can carry back into later conversations.
The change addresses a basic continuity problem: users should not have to explain the same conference, team or project each time they switch between asking a question and delegating work. Anthropic is making the feature available by default to Free, Pro and Max users across web, desktop and mobile, although mobile users need the latest app version.
One memory layer across two modes
Claude chat is built around conversation, while Cowork handles more extended work. Under the new arrangement, both can draw from the same retained knowledge. Anthropic’s example is conference planning: after learning a city’s location, expected attendance and speaker details in chat, Cowork could use that context when helping with related tasks.
Memory is added during the conversation rather than being assembled only as an end-of-chat summary. Anthropic describes the effect as a more continuous assistant that can accumulate context as work develops. The source does not provide measurements showing how often the system recalls the right fact or avoids applying it to the wrong project.
Users can inspect, edit and delete what Claude has retained. Those controls are particularly important when a remembered detail becomes outdated or was saved without the context that originally made it accurate. Shared memory reduces repetition, but it also increases the consequences of a mistaken entry because the same error can surface in more than one mode.
Sensitive information has separate boundaries
Anthropic says Claude will not save personal or sensitive details by default. Its examples include health information, race, ethnicity, religion, political views and gender identity. A user can opt in to retaining such material through a separate toggle, and Claude will display a notification when it saves a sensitive detail.
Some categories are excluded even with that choice. Anthropic says the system will never save government identification, Social Security numbers, criminal history, immigration status or information prohibited by its acceptable-use policy.
These are provider rules for the memory product, not an independent privacy audit. The source does not explain how reliably the system classifies ambiguous information or how long deleted entries persist in underlying systems. People using Claude for consequential work will still need to review what has been stored and avoid treating the assistant as the sole record of a sensitive project.
Continuity creates both leverage and risk
A shared memory can make an agent more useful because context becomes available at the point of action. Cowork may not need a new briefing before every assignment, and chat can discuss the results without reconstructing the background. That is a practical shift from isolated sessions toward an assistant that follows work over time.
The same continuity can blur boundaries. Details appropriate for one task may be irrelevant or confidential in another, while a preference learned casually can become an assumption in future work. Visibility and deletion controls help, but users must notice when remembered context is influencing an answer.
Anthropic’s rollout makes persistent memory a default part of Claude for a broad user base rather than an advanced organizational feature. It also gives people an explicit view of stored topics and separate consent for sensitive categories. The decisive test will be whether the system can provide useful continuity without surprising users about what it remembers or where that information appears.
For now, the update unifies two previously separated experiences. Claude is becoming less like a collection of independent chats and more like an ongoing workspace with retained context. That can remove friction from repeated work, provided the memory remains accurate, understandable and under the user’s control.
