Claude can now carry the same memory between chat and Cowork, which saves repetition and raises a simple question: what exactly did you let it remember?
Anthropic says Claude’s memory now works across regular conversations and cloud Cowork tasks. The continuity is useful when a delegated job needs your preferences, project context, or working style. The default is not the same for everyone, though. Memory is on by default for Free, Pro, and Max accounts, and off by default for Team and Enterprise.
Chat and Cowork now share one memory
Anthropic’s August 25 announcement says the same memory is available in chat and cloud Cowork. A preference learned in conversation can therefore affect a delegated Cowork task, and context from work can remain available when the user returns to chat.
This is product memory, not proof that an agent completed the work correctly. Our article on PAST-Bench and agent memory explains the harder test: stored context should improve a later task without introducing stale or misleading instructions.
The default depends on the plan
| Account type | Documented default | First check |
|---|---|---|
| Free | Memory on | Review stored topics before sensitive work |
| Pro | Memory on | Separate personal and client context |
| Max | Memory on | Audit long-running project assumptions |
| Team | Memory off | Confirm workspace policy before enabling |
| Enterprise | Memory off | Define retention and review ownership |
The consumer default optimizes for continuity. The business default optimizes for a deliberate rollout. Neither is automatically correct for every task. A freelancer handling several clients may need stricter separation than a single personal account provides. A company may decide that certain non-sensitive preferences are worth enabling after policy review.
Users can inspect and delete memory topics
Anthropic says users can open Settings, select Memory, and read, edit, or delete saved topics. That is the control I would check before the first Cowork delegation, not after a surprising answer. A memory entry can be accurate when written and wrong a month later.
Start with preferences that age slowly: preferred output format, recurring project vocabulary, and stable communication style. Be cautious with temporary deadlines, current prices, client status, access rules, and anything that should expire. Memory without an expiration habit becomes a stale configuration file.
Anthropic excludes some sensitive categories
Anthropic says Claude excludes sensitive subjects from memory by default and lets users opt in to some categories. It also says some information is never stored, including government identification numbers, criminal history, and immigration status.
That product boundary is useful, but it should not become a data-classification strategy. Do not paste a sensitive record merely because the memory layer says it will not retain one field. The conversation, connected tools, workspace policy, and exported output still need their own controls.
Run a memory handoff test before real delegation
- Create a harmless test project with one stable preference and one temporary fact.
- Ask Claude in chat to summarize what it should remember.
- Open Settings and inspect the stored topic.
- Run a Cowork task that should use the stable preference.
- Change the temporary fact, then confirm the next task does not reuse the old value.
- Delete the test topic and verify that continuity disappears.
That test reveals whether your team understands the difference between conversation context, saved memory, project files, and live tool data. Those sources can disagree. A good workflow says which one wins.
Memory should reduce setup, not review
The broader Claude comparison guide helps choose the assistant. This update changes the operating habit after that choice. You may spend less time restating preferences, but the final work still needs a fresh evidence and acceptance check.
Cross-product memory is most valuable for stable context and most dangerous when a temporary assumption quietly becomes permanent. I would keep it on for personal formatting preferences, review it monthly, and keep client secrets and fast-changing operational facts out of the memory layer.
My verdict
Claude memory across chat and Cowork is a real convenience, not an excuse to delegate blindly. Check the default, inspect the stored topics, and prove that deletion and correction work before a serious project depends on them. Continuity earns trust only when the user can see and change what survives.
Read Anthropic’s memory announcement and its current Claude release notes before changing a workspace policy.
Memory defaults and controls were verified from Anthropic’s public documentation on September 1, 2026.
