Anthropic integrates Claude chat with Cowork memory, which is enabled by default.

      Anything you communicate to Claude in a chat window is now accessible to Claude Cowork. Knowledge gained by Cowork is returned to Claude as well. On Tuesday, Anthropic merged both memory systems. As a result, whenever you interact with Claude, it builds on what it already knows about you, according to the company's announcement.

      This feature is enabled by default for consumers across Free, Pro, and Max plans, available on web, desktop, and mobile platforms.

      The second significant change is that Claude now adds topics to memory while you're still conversing. Previously, it would wait until the end of a conversation to summarize it. For instance, if you indicate that a deadline has moved to September, the subsequent conversation will be aware of this change without you needing to say, “remember this.” This alters how data is recorded, effectively eliminating the pause at the end of a conversation, which was the moment when a user might have opted not to keep the exchange.

      In practice, if you ask Cowork to draft an update for your manager, it will recognize who that is and how she prefers updates to be written—this is an example provided by Anthropic. If you brainstorm a conference agenda in the chat, Cowork will retain knowledge of the headcount, city, and speakers while it prepares the budget document.

      Sarah Perez succinctly expressed in TechCrunch that this update makes Claude feel more like a cohesive assistant rather than two separate products under one umbrella.

      What is stored can be reviewed; everything Claude remembers is arranged in a list of files under Topics in the memory settings. Users can read, modify, or delete each entry, and any corrections will apply across the board. At any time, users can pause or reset memory, with two toggles available—one for generating memory from chats and another for sensitive topics.

      Regarding sensitive topics, by default, Claude does not keep any information on health, race, ethnicity, religious beliefs, politics, or gender identity. The company has outlined these categories and included others that are similar. Users have the option to activate these settings. Anthropic argues that while some may view certain topics as sensitive, others may find them useful; for example, a gluten allergy that could influence recipe suggestions. The switch remains off unless actively turned on by the user, and it does not apply retroactively. Each time Claude saves something in one of those categories, it displays a notification.

      David Gewirtz pointed out a potential concern at ZDNET. For individuals seeking help with sensitive issues, receiving a warning every time they discuss it may not be comforting.

      There are certain categories that remain off-limits even with the sensitive switch activated. This includes social security and government identification numbers, criminal history, immigration status, and anything that contravenes Anthropic’s acceptable use policy. Claude will inform users when it cannot update memory with such information.

      Claude Code is not included in this update. Gewirtz inquired and was informed that the focus is solely on Cowork and chat, leaving the coding tool as an isolated memory space, indicating that this consolidation is still ongoing.

      The scope of the service is broader than just two products. Cowork, which Anthropic introduced at the beginning of the year, has expanded since its launch on phones in July. Igor Bonifacic highlighted the timing for Engadget; a recent update integrated Cowork into the Claude for the Chrome extension, meaning conversations in the browser's side panel are now part of a user's activity history.

      Introducing retention in browsers can be sensitive, especially since Chrome has faced criticism for silently installing features without user consent. According to Bonifacic, Claude has been capable of retaining memories since last summer, though those memories remained segmented partly by choice and partly due to the system's architecture.

      The significance of this reach lies in the fact that Cowork is not merely a chatbot; it performs tasks, including cloud-based functions, and operates on a user's device. In July, researchers discovered that Claude Cowork could access login credentials on a Mac by escaping its local virtual machine.

      This makes the memory aspect distinct from standard chat interactions. A system that can take action and now possesses months of accumulated user context is a more formidable entity than either of its individual components.

      As for defaults, memory is automatically activated for consumers, while Team and Enterprise customers do not have this feature enabled by default. Administrators control organizational settings, and memory remains inactive for individual users until they opt to turn it on. Gewirtz reported the enterprise situation slightly differently, indicating that organizations have this feature off by default, subject to eligibility criteria. Regardless, this division is intentional, placing the responsibility of the decision on compliance teams rather than individual employees.

      This distinction is particularly significant in Europe, where companies must document what personal data their tools retain. A recent fine of 825 million euros for Uber over automated decision-making serves as a reminder of the regulatory climate rather than a critique of this