On August 25, Anthropic announced that Claude Chat and Cloud Claude Cowork now share the same memory system.

Article author and source: AI New Era

You just gave it a task, but it even remembers your boss.

In the future, hand the task over to Claude and don't repeat the introduction.

The project you discussed in chat for three months: Who are the clients? What does the boss want? How many columns should the report table have? Which number cannot be included in the conclusion…

Everything you accumulate in Chat, Cowork will remember when you hand over the task.

New situations that arise while Cowork is working will be sent back to Chat via the same path.

Previously, Chat and Cowork were like two separate individuals.

In chat, you get familiar with someone; in Cowork, you have to reintroduce yourself every time—starting from “Who I am” and “How our team measures metrics.”

The cold start between chatting and working has now been canceled.

Memory has shifted from a feature in the chat box to the context layer during agent operations.

On August 25, Anthropic officially announced: Claude Chat and Cloud Claude Cowork will now share the same memory system starting today.

Claude's memory is now everywhere, and you decide what it remembers.

Boris Cherny, the creator of Claude Code, shared this announcement, calling it a small improvement: memory is now simpler and more powerful.

Say it once in chat, no need to repeat when working.

This update is easier to understand in everyday scenarios.

Have Cowork send a work update to the manager; it knows who the manager is and what style she prefers.

Brainstorm an agenda for the meeting in the chat box, then let Cowork handle the budget and execution documents—it doesn’t need you to repeat the number of attendees, city, or speaker list.

In the chat, clearly explain once how your team defines the metrics; thereafter, Cowork will apply this same set of metrics to every quarterly business review document, eliminating the need for repeated briefings.

Moreover, the shared memory will be continuously updated.

New contexts that appear in Cowork tasks will flow back into Chat.

The timing is also interesting.

On August 18, Cowork rolled out its paid plans on the web, desktop, and mobile platforms. Seven days later, memory was unified.

This update didn’t focus on enhancing the model’s capabilities, but rather on how much it knows about you before getting started.

Understanding you helps me serve you better.

The black box of memory has become a transparent ledger that can be modified at any time.

When people see "Memory Unlocked," their first thought is: "My chat history has been moved."

That's not how it is.

The official help center clearly states: Claude stores a set of entries generated by topic, called Topics.

And it saves as you chat, rather than waiting until the end of a conversation to summarize everything at once.

What's the difference?

You casually mention that the project deadline has been moved to September, and when you start a new conversation, it already knows—no need for you to restate it. Of course, if you're unsure and want to remind it again, you can, and it will keep the note.

More importantly, you can see and edit these items.

Go to Settings > Memory, where you’ll find all the content Claude remembers listed under Topics. Tap any item to view, edit, or delete it.

The Topics list on the Memory settings page contains all entries remembered by Claude—readable, editable, and deletable.

Because these files are all short, they are easy for you to maintain.

For example, if the company changes its name, update that entry, and ensure the change is reflected in every subsequent conversation.

If you don't want it to remember, there are two options.

Pause: Existing memories will be retained but no longer used or added to; conversations during the pause will not be logged. Reset: Delete everything, including project memories—this action is irreversible.

Moreover, each project has its own independent memory space and project summary, with no interference between them.

Previously, AI remembered those things about you, and you could only guess from its responses.

Now, your profile information is a document you can edit anytime from the settings page.

The memory function has changed from a black box to a transparent ledger that can be modified at any time.

Three-layer memory switch

A major design feature this time is Anthropic’s three-tiered switch for memory.

Layer 1, default storage.

Your role, projects, work and life contacts and locations, communication preferences, technical preferences, coding style, and ongoing tasks... This layer typically includes the context needed for daily collaboration.

Layer 2 is not stored by default, but you can enable it yourself.

Health, race, ethnicity, religious beliefs, gender identity, and similar categories.

Officially, some people find it sensitive, while others find it useful.

For example, when you turn this on, Claude will remember your gluten allergy when planning your weekly meals.

This switch is governed by three detailed rules.

Applies only to content after activation; no retroactive effect on prior content.

Each time you submit this type of content, a prompt will appear above the input field, allowing you to review or adjust your settings immediately.

If you decline the prompt or later turn off the switch, the sensitive entries already saved will be cleared.

Layer three: never stored.

Identification number, criminal record, financial account number, immigration status, and any content violating the usage policy.

This layer does not have a switch.

Even if you explicitly ask Claude to remember it, it won’t store the information and will clearly explain why it can’t.

Have Chat and Cowork been integrated?

Not all of them—only Cowork in the cloud has been integrated.

The official statement appears in both the blog and the help center: This shared memory only takes effect when Cowork is running in the cloud.

Excludes local sessions running on your own computer.

Can free users also use the unified memory workflow?

No.

By default, memory is enabled for Claude at the Free, Pro, and Max tiers, not Cowork.

Cowork is still only available to Pro, Max, Team, and Enterprise users.

Different terminals have different scenarios.

The macOS and Windows desktop versions are available for all paid plans. The web and mobile versions are available for Pro, Max, and Team; Enterprise is subject to administrator enablement.

Team and Enterprise are disabled by default and must be enabled by the organization owner in the settings.

There’s another hidden risk here: once the owner turns off organizational memory, all memory entries for everyone in the organization will be permanently deleted immediately.

Additionally, organizations that have signed HIPAA, public sector, or certain custom data retention agreements cannot currently use Memory.

The conversation has expired or been deleted, but the memory entries generated from it will not be deleted.

To clear it, you need to delete it individually in Settings > Memory.

This is the one thing most ordinary users should know—and the most likely to overlook.

After connecting, can I still discuss sensitive topics with Claude?

Just chat.

Only the action of "saving to long-term memory" is disabled by default, not the chat itself.

You want to talk about health—Claude will listen to every word and chat with you fully, but won’t store anything in memory afterward.

The next round of competition is about who doesn't need to get to know you again.

What is Cowork?

The official definition is: applying Claude Code’s agent architecture to knowledge work beyond programming, without requiring a terminal.

It can break down tasks, parallelize sub-agents, run continuously in the cloud, keep running even when you close your laptop, and can be scheduled as a cron job that executes even when your computer is powered off.

Now, attach to it an entire set of context that has been accumulated over several months.

Behind this small improvement lies a new change.

Model capabilities can catch up—today you're two months ahead, but tomorrow your competitor may close the gap. However, the context accumulated by users is much harder to match.

Replacing a system that already knows how you write your weekly reports, how your team measures metrics, and who you worked with last quarter isn’t just about the subscription cost—it’s about having to explain yourself all over again.

Interestingly, Anthropic left the door open.

Official support for importing and exporting memories between Claude and other AI services is available, though still marked as experimental. The door is open, betting that you won’t leave—but others will come.

In the previous round of agent competition, everyone was competing for tasks: who could complete the work and who could sustain it longer.

What’s at stake in the next round of competition has changed: it’s your memory—the things you don’t want to explain a second time—and ultimately, where they’re stored.