Claude's Memory Now Works Across Both Chats and Cowork Sessions
Anthropic has updated Claude's memory feature so that stored memories are accessible across both the Claude chatbot and Claude Cowork, the company's computer-use agent released at the start of 2025 and generally available on macOS and Windows since April 9, 2026. When Cowork runs a task in the cloud, what Claude has learned about the user from chat conversations is available to it, and conversely, what Cowork encounters during task execution feeds back into the shared memory store (Engadget).
The underlying memory mechanism is file-based. Claude stores and retrieves information across conversations in a directory of memory files, and the model can create, read, update, and delete those files on its own (Anthropic docs). Users can also manually add specific information into memory via Settings > Capabilities and clicking "View and edit your memory" (Anthropic support). Memory is enabled by default for all Free, Pro, and Max accounts (Engadget).
Alongside the cross-surface unification, Anthropic added a new "include sensitive topics in memory" toggle to Claude's memory controls. By default, Claude does not store topics related to personal or sensitive subject matter, explicitly including health, race, ethnicity, religious beliefs, politics, and gender identity. When the toggle is enabled, Claude notifies the user each time it adds such a topic to memory. Users can turn the feature off at any time and can edit or delete individual memory entries (Engadget).
Anthropic has drawn hard lines around certain data categories regardless of the toggle state. Claude will not store social security numbers, criminal history, or immigration status even when the sensitive-topics option is enabled. The company also states that Claude will not retroactively add past conversations to memory; only new interactions going forward contribute to the memory store (Engadget).
For Team and Enterprise plans, the interaction between memory and Projects adds another management layer. Pro and Team users can organize chats into Projects, which consolidate internal knowledge and chat activity in one place (Anthropic). The ability to move chats into and out of Projects effectively lets users control what falls within Claude's memory scope on those plans (Anthropic support).
The cross-surface memory update arrives amid a busy stretch for Anthropic. The company introduced Claude Opus 5 on July 24, 2026, describing it as a step-change improvement for the Opus tier optimized for long-running agents, coding, and professional work. Claude Sonnet 5 followed on June 30, 2026, focused on frontier performance across coding, agents, and professional work at scale. On August 14, Anthropic published technical details on Claude's text watermark, and on August 7 it announced improvements to Claude Fable 5's biology safeguards to reduce false positives. The company also published its position on open-weights models on July 27 and detailed three real-world cybersecurity incidents from its evaluations on July 30 (Anthropic newsroom).
The practical significance of unifying memory across chat and Cowork is straightforward: it eliminates the context barrier between conversational interaction and autonomous task execution. A user who refines a project specification through back-and-forth chat can hand off execution to Cowork without re-establishing preferences, constraints, or domain context. The agent picks up what the chatbot already knows.
The broader context here is a design tension that anyone building persistent memory into a consumer-facing AI assistant has to navigate. Persistent memory that spans surfaces is genuinely useful for productivity; it is also a concentrated store of personal information that now flows from chat into cloud-based agent execution. Default-off for sensitive categories, per-event notifications when the toggle is on, and categorical exclusions for identifiers like SSNs and immigration status are Anthropic's attempt to give users granular consent without forcing them to choose between utility and privacy.
Worth noting is that the file-based memory architecture means Claude can create, read, update, and delete memory files autonomously. For users on Team and Enterprise plans, the Projects mechanism provides a structural way to scope what the model can see, but individual users on Free, Pro, and Max tiers are relying on per-entry editing and deletion for ongoing memory hygiene. The retroactivity exclusion is a meaningful guardrail: it means the memory store only grows from the point of activation forward, not from the full history of prior conversations.
For developers building on Claude's API or integrating Cowork into workflows, the shared memory layer means that context established in one surface is no longer siloed. Whether that unification reduces friction enough to shift agentic workflow patterns will depend on how reliably the model retrieves and applies the right memories at the right moments. The mechanism is in place; the quality of retrieval is what will determine its practical value.
