Claude Memory Now Works Across Chat and Cloud Cowork: What Carries Over—and What Doesn't

Short answer: yes—but the important word is cloud. On August 25, 2026, Anthropic said Claude’s memory now works across regular chat and Cowork sessions that run in the cloud. What Claude learns in chat can help with a cloud Cowork task, and useful context from that Cowork task can carry back into later chats.

That does not mean every Claude surface now shares one giant, unrestricted memory. Projects keep separate memory spaces, local Cowork sessions do not share this cross-surface memory, and Incognito chats neither read from nor write to memory. Claude also gives users a list of remembered Topics that can be edited or deleted.

The practical change is bigger than “Claude remembers more.” It changes how to choose between a normal chat, a Project, Cowork, and Incognito mode depending on whether context should travel with the work.

Feature check — August 29, 2026: Anthropic’s current release notes and Help Center say memory is on by default for Free, Pro, and Max accounts. Team and Enterprise organizations have memory off by default unless enabled by an owner. Cloud Cowork is currently available on paid plans, with web and mobile in beta.

The simplest mental model: four different context boundaries

The easiest way to avoid surprises is to stop thinking of “Claude memory” as one feature and instead think about where context is allowed to travel.

This distinction matters because two conversations that look similar in the interface can have very different persistence rules.

Anthropic’s Help Center says each Project has its own memory and project summary, separate from other Projects and from non-project chats. So a useful piece of context learned inside “Client A” should not automatically become part of a separate “Client B” Project’s memory.

Meanwhile, Anthropic explicitly says the new Chat↔Cowork memory connection applies when Cowork runs in the cloud. The Help Center notes that it is not available for Cowork sessions that run locally on a computer.

What actually changed on August 25?

Before this update, Claude’s memory and Cowork could feel like adjacent systems: chat could remember conversational context, while Cowork handled longer-running agentic work.

Anthropic’s August 25 release adds three practical changes.

1. Chat memory can follow a task into cloud Cowork

If Claude already knows recurring context—such as a project name, a preferred writing style, or a technical convention—a cloud Cowork task can use that memory without requiring the same background to be pasted again.

Anthropic gives a simple example: a Cowork task can draft an update for a manager while already knowing who that person is and how they prefer updates written.

The useful implication is broader: memory can now reduce setup cost for agentic work, not just make conversations feel more continuous.

2. Cowork can contribute context back to later chats

The flow works in both directions. Relevant context that comes up during a cloud Cowork task can become available when returning to regular Claude chat.

That means a multi-step task no longer has to end as an isolated execution session. If Cowork learns a durable preference or project fact that Claude decides belongs in memory, later chat can use it.

3. Memory is now exposed as editable Topics

Anthropic says everything Claude remembers in the new experience appears under Settings → Memory → Topics. Individual items can be edited or deleted.

This is an important design change because persistent AI context is much easier to trust when it is inspectable. Instead of treating memory as a mysterious summary somewhere in the system, users can audit what is being carried forward.

The biggest gotcha: deleting a chat is not the same as deleting its memory

This is the part worth checking before assuming conversation deletion resets everything.

Anthropic says that when a conversation expires or is deleted, related memory entries generated from that conversation are not automatically removed. Those memory items can still be deleted separately from the Memory settings.

So there are now two different cleanup actions:

1. Delete the conversation when the chat itself should disappear from history.
2. Delete the related memory Topic when the learned fact or preference should stop carrying forward.

For a one-off embarrassing typo, this distinction may not matter. For outdated project details, changed preferences, or a mistaken fact Claude stored, it matters a lot.

A good habit is to review Topics after any major context change rather than assuming old information will disappear when old chats are removed.

Sensitive topics are opt-in, not normal memory

Anthropic says Claude’s memory excludes categories of sensitive personal information by default, including topics such as health, race, ethnicity, religious beliefs, politics, and gender identity.

A separate Include sensitive topics in memory setting can be enabled for people who deliberately want Claude to retain some of that context. Anthropic says the setting applies going forward rather than retroactively.

The company also says some information is not saved to memory even when requested, including government ID numbers, criminal history, financial account numbers, and immigration status.

The practical rule is straightforward: do not treat ordinary memory as a secure vault for secrets or identifiers. Its purpose is continuity, not credential storage.

Incognito is the “do not carry this forward” mode—but it is not zero-retention mode

Claude’s Incognito chats are the cleanest option when a conversation should not influence future memory.

According to Anthropic, Incognito chats:

• do not use Claude’s existing memory;
• do not create future memory entries;
• are not saved to normal chat history;
• are not used when Claude searches previous conversations;
• are available on Free, Pro, Max, Team, and Enterprise plans.

But Incognito does not mean the conversation instantly ceases to exist on Anthropic’s systems. Anthropic says Incognito chats are retained for 30 days by default, or longer where an Enterprise organization has configured a different data-retention policy.

That makes the distinction useful:

Memory controls what Claude carries into future work. Data retention controls how long a conversation may remain in the provider’s systems. They are not the same thing.

For Team and Enterprise users, Anthropic also says Incognito conversations can still be included in organizational exports and, for Enterprise, the Compliance API.

Pause memory or reset it? They solve different problems

Claude currently offers two very different ways to stop memory.

Pause memory

Pausing means Claude keeps the memories that already exist but stops using them and stops creating new memories. Anthropic says conversations that happen while memory is paused are not added later if memory is turned back on.

Use Pause when the goal is temporary separation without destroying the existing memory set.

Reset memory

Resetting permanently deletes memory, including Project memories. Anthropic says this cannot be undone.

Use Reset only when the desired outcome is genuinely to start over.

That leads to a useful three-level decision:

Deleting a single Topic is the more surgical option when only one fact is wrong or outdated.

Projects are still the better boundary for work that should stay separate

Cross-chat memory can be convenient, but convenience becomes a problem when unrelated work starts influencing the same context pool.

Projects remain the better structure when context should be persistent but compartmentalized.

A practical setup could look like this:

• Regular chat: general preferences and everyday context.
• Project A: a product launch, with its own files, instructions, and memory.
• Project B: a separate client or research topic that should not inherit Project A’s details.
• Cloud Cowork inside a Project: longer-running tasks that should use that Project’s established context.
• Incognito: one-off conversations that should not become memory at all.

This is more reliable than asking Claude to “forget about the other project” inside a shared conversation. The product boundary itself does more of the separation work.

Does this cost extra?

Memory itself is not a paid-only feature. Anthropic’s current pricing page lists memory across conversations on the Free plan.

Cowork is different: it is a paid-plan capability. As checked on August 29, 2026:

On Team and Enterprise, Anthropic’s August 25 release note says memory is off by default at the organization level. On Enterprise, Cowork availability and cloud execution can also depend on admin settings.

So someone on Free can use Claude memory across ordinary chats and Projects, but the specific chat↔cloud Cowork workflow requires access to Cowork.

A five-minute memory audit

The new cross-surface behavior makes this a good time to review Claude’s settings before relying on memory for important work.

1. Open Settings → Memory → Topics

Scan for facts that are outdated, overly broad, or simply wrong.

A persistent incorrect assumption is worse than having no memory because it can quietly influence many future sessions.

2. Decide what belongs in a Project

If a piece of context is specific to one client, product, research topic, or workstream, move the work into a Project rather than allowing it to live as general account memory.

3. Check the sensitive-topics setting

Leave it off unless there is a clear reason to let Claude retain those categories between conversations.

4. Use Incognito intentionally

Incognito is useful for exploratory questions, unusual one-off tasks, or conversations that should not change Claude’s future assumptions.

Remember that it is a memory/history boundary, not an immediate-retention guarantee.

5. Test Cowork with something harmless first

Before depending on cross-surface memory for an important workflow, run a low-risk test:

1. Tell Claude a non-sensitive preference in normal chat.
2. Check that it appears appropriately in Memory Topics.
3. Start a cloud Cowork task where that preference is relevant.
4. Confirm Claude uses it correctly rather than merely assuming it will.
5. Return to chat and verify that any useful Cowork-derived context is reflected accurately.

This distinguishes the documented feature from the quality of memory retrieval in a particular workflow.

What this update does not mean

Claude does not remember every sentence forever

Anthropic describes memory as categorized entries that Claude reads and updates, not a verbatim permanent transcript of every conversation.

Every Project does not share one memory

Projects have separate memory spaces specifically to keep their context focused.

Local Cowork is not the same as cloud Cowork

The new memory bridge is documented for Cowork in the cloud. Local Cowork sessions are explicitly excluded from that cross-surface memory behavior.

Incognito does not mean “not retained anywhere”

It prevents normal history and memory use, while Anthropic separately documents a default 30-day retention period.

Turning memory off does not automatically delete it

Pause and Reset are separate choices. Pausing stops use and creation; resetting deletes the memory set.

What to watch next

Three areas are worth watching as Anthropic expands this system.

More consistent organization controls

Team and Enterprise memory depends on owner/admin configuration, and Anthropic is still rolling out newer memory experiences across account types. Organizations will need clear policies for which persistent context is useful versus inappropriate.

Better visibility into why a memory was used

Editable Topics solve the “what is stored?” question. The next useful layer would be clearer explanations of which memory influenced a specific answer or task—especially when agents act across files, browsers, and connected services.

Memory quality, not just memory quantity

The useful benchmark is not how much Claude can remember. It is whether Claude stores the right durable facts, retrieves them at the right time, and makes it easy to correct stale information.

For regular users, that means occasional Topic reviews. For teams, it means treating AI memory more like shared configuration or knowledge management than a magical personal notebook.

Conclusion

Claude’s August 25 memory update makes persistent context substantially more useful because it can now follow work from chat into cloud Cowork and back again.

But the feature is easiest to use safely when its boundaries are explicit: Projects isolate ongoing work, Incognito blocks memory for one-off conversations, local Cowork remains outside the new bridge, and deleting a chat does not automatically delete memories derived from it.

The best setup is not “turn memory on and forget about it.” It is to decide which context should be general, which belongs inside a Project, which should travel into agentic tasks, and which should never persist in the first place.

Sources

Checked August 29, 2026:

• Anthropic Help Center — Release notes
• Anthropic Help Center — Use Claude’s chat search and memory to build on previous context
• Anthropic Help Center — Use Claude Cowork on web, desktop, and mobile
• Anthropic Help Center — Use incognito chats
• Claude — Plans & Pricing