Claude Cowork now carries chat memory into tasks

Claude Cowork now carries memory from Claude chat into tasks, giving Anthropic's agent a shared context instead of two separate briefing loops. Anthropic announced the change on August 25, 2026, and the rollout also adds topic-level controls so users can inspect, edit, or delete what Claude retained. For anyone using an AI agent for multi-step work, the practical change is simple: explain a project once, then move between discussion and execution with less re-entry.

Definition: Claude Cowork and Claude chat now use one shared memory when Cowork runs in the cloud.

Example: A Cowork task can use details previously discussed in chat, such as a manager's preferences or a conference's headcount and speakers.

Key takeaway: Claude's memory is moving from a chat-only convenience to context shared across an agent workflow.

Business impact: Teams can hand work from research and brainstorming to execution with fewer repeated instructions, but they still need to review what the system stores and where the task runs.

What changed in Claude Cowork?

Anthropic merged the memory systems used by Claude chat and Claude Cowork, so Claude's official release notes now describe memory as working across chat and Cowork in the cloud. The feature is meant for the handoff between thinking and doing: a user can discuss a project in chat, then ask Cowork to act on the same project without reconstructing the background. The immediate action for a team is to treat shared memory as context to verify, not as a replacement for a clear task brief.

Before this update, Claude Cowork could force users to repeat details already established in chat when they switched from conversation to an agent task. The new design makes Claude feel more like one continuous assistant across those surfaces, according to TechCrunch's report on the announcement. The useful shift is not that Cowork can perform a new class of task; it is that the same project context can follow the user into the task that needs it.

Why does shared memory matter for agent workflows?

Shared memory matters because an AI agent usually needs both a goal and the background that makes the goal specific. Claude Cowork can now receive a task after a user has discussed the project in chat, so a request such as drafting a manager update can retain relevant preferences instead of starting from a blank context. The practical takeaway is to separate stable project facts from one-off instructions and check that the retained facts are still correct before an agent acts.

Claude Cowork can carry details encountered during a cloud task back into Claude chat, so the two surfaces operate as stages of one workflow rather than separate products under one account. This handoff extends the cross-device Cowork workflow covered in Yowox's earlier report. Businesses evaluating agent adoption should therefore test the handoff between planning and execution, not only the quality of either surface in isolation.

What does Claude remember during a conversation?

Claude now adds memory as topics emerge during a conversation instead of waiting until the conversation ends to summarize it. Anthropic's memory documentation says the system can remember everyday context such as a user's role, projects, communication preferences, technical preferences, and ongoing work. The operational takeaway is to review fast-changing project details, because a memory system that updates during a chat can carry a changed deadline or preference into the next task sooner.

Claude's memory model differs from searching an old transcript: Claude's chat search can retrieve relevant past conversations, while Claude memory stores selected topic-level context for new conversations and cloud Cowork tasks. The distinction matters for operators, who should use search when the exact history matters and memory for durable context that should shape future work.

Which controls do users get?

Claude exposes remembered information as individual topics under Settings > Memory, where users can read, edit, or delete entries. Users can also ask Claude in a conversation to remember something, change a memory, or forget it, and the change applies to future conversations. The concrete safeguard is simple: inspect the topics before relying on them for a recurring workflow, and delete or correct entries that no longer describe the work.

Claude's sensitive-topic setting excludes health, race, ethnicity, religious beliefs, politics, and gender identity by default. Users can opt in through an “include sensitive topics in memory” setting, and Claude notifies them when a sensitive topic is saved. The setting is not a universal data vault: government-issued identification, Social Security numbers, criminal history, and information that violates Anthropic's acceptable-use policy remain excluded.

Who gets Claude's memory feature?

Claude memory is enabled by default for Free, Pro, and Max plans across Claude's web, desktop, and mobile experiences, while Anthropic's support documentation says Team and Enterprise organizations have memory off by default unless an administrator enables it. Mobile users need the latest iOS or Android app for the update. Teams should confirm their plan and app version before relying on the rollout.

Where does Claude's shared memory work?

Claude's shared memory works when Cowork runs in the cloud, not when a Cowork session runs locally on a computer. A cloud Cowork task can use context from Claude chat and return context to Claude chat, while a local Cowork run does not participate in that cross-surface memory. Teams should confirm the execution mode before treating Claude memory as a universal layer.

What should teams watch next?

Claude Cowork's shared memory removes one visible source of friction in agent workflows: re-explaining the same project every time work moves from discussion to action. The feature also increases the importance of memory review, because an incorrect preference or stale project fact can now travel into a task more easily. Teams should start with low-risk workflows, inspect the topic controls, and keep sensitive or high-impact decisions in an explicit brief that a human can review.

The broader AI-agent market signal is that products are competing on continuity as much as on individual task performance. Claude Cowork's update does not prove that memory will always select the right context, and it does not make local sessions equivalent to cloud sessions. It does show where Claude Cowork is heading: a user describes intent in chat, an agent acts on it in Cowork, and both surfaces retain enough context to continue the work.