Anthropic unifies Claude memory across Chat and Cowork
Anthropic has updated Claude’s memory system so information can persist across Claude Chat and Cowork, while adding finer controls for sensitive subjects and topic-by-topic review. The change is designed to reduce the need for users to repeat preferences and project context when moving between conversational work and Cowork’s multi-step agent workflows.\n\nThe update gives users a dedicated Memory area in Claude’s settings, where remembered information is organized by topic and can be reviewed, edited, or deleted. Anthropic is also adding a control that lets users decide whether Claude may retain sensitive topics. Sensitive information is excluded by default, according to reporting on the rollout, with the setting intended to make inclusion an explicit choice.\n\nThe cross-product change matters because Chat and Cowork serve different modes of work. Claude Chat is used for questions, drafting, and brainstorming, while Cowork can perform multi-step knowledge-work tasks such as creating documents, handling files, and carrying out agentic workflows. A shared memory layer can carry user preferences and ongoing context between those modes, potentially making transitions less repetitive.\n\nAnthropic previously described memory as a user-visible summary that people can inspect and modify, with project-scoped memory helping keep separate areas of work from mixing. Its Cowork documentation also describes projects as containing files, instructions, links, and memory. The latest rollout extends the idea of continuity beyond a single project or interaction path, while making topic boundaries more visible.\n\nThe main benefit for users is control paired with continuity: people can see what Claude has retained, remove outdated information, and make a deliberate decision about sensitive categories. The trade-off is that a broader memory surface increases the importance of reviewing stored context, especially when Cowork has access to files or connected tools. Existing Cowork guidance warns users to review planned actions and consider the sensitivity of files they authorize.\n\nThe update does not come with a public benchmark, price change, or quantified performance improvement in the available reporting. It is a product and privacy-control change rather than a new foundation-model release. Availability and plan coverage may vary by platform and account as the rollout proceeds, and Anthropic’s help documentation has contained transitional limitations around memory support in Cowork, so users should check the current controls in their own Claude settings.\n\nFor developers and knowledge workers, the significance is less about raw model capability than about persistent context becoming a shared product layer. The practical test will be whether topic-level controls keep memory useful without allowing stale, overly personal, or unintended information to influence future agent actions.