Every time I handed a task to Cowork, I started explaining again. How I work, which two brands have to stay apart, why a Dutch text has a different rhythm than an English one. I had already said all of it in chat. One half simply had no idea what I had told the other.
Since 25 August that is one memory. Anthropic announced it on its own blog, and product manager Amol Avasare summed it up on X as “memory is now merged between Claude chat and Cowork”. Claude has remembered things for a while now. What changed is that where it remembers no longer depends on which surface you happen to be in.
What was actually announced
Three things at once, and they belong together.
| Part | What it does | Where it applies |
|---|---|---|
| Shared memory | What Claude knows from your chats is there when you hand Cowork a task, and the other way round | Only when Cowork runs in the cloud |
| Memory during the conversation | Claude files topics while you talk, instead of summarizing afterwards | Everywhere memory is on |
| Sensitive topics | Health, beliefs, politics and similar categories stay out unless you switch them on | Off by default |
The middle one changes the most. Until now Claude distilled a conversation only once it was over. Anthropic now writes: “Claude now adds topics to memory as you chat, instead of summarizing conversations after they end.” That is the difference between a diary written up at night and a notepad that stays open all day. For moving back and forth between chat and Cowork it is exactly what is needed, because otherwise an ongoing conversation would have handed over nothing yet.
The memory itself is no longer a black box. Everything sits as a list of short files under Topics in Settings > Memory, and you can read, edit or delete each one. Anthropic gives a reason I recognize from my own work. Correct your company’s old name in one file and every conversation after that gets it right.
The limit that is missing from the announcement
There is tension between the sources here, and I am not going to resolve it by picking one.
The blog post says the memory works everywhere. The Help Center adds something the blog post never mentions: “Memory across Cowork and chat only works when Cowork runs in the cloud. It isn’t available in Cowork sessions that run locally on your computer.” The release notes back that up, because they say “across chat and Cowork in the cloud”, with those last three words attached.
That is not a footnote for me. I wrote earlier about Cowork running on without my laptop and how that rearranges a working day. The flip side only becomes visible now. The sessions you deliberately keep local, and you usually keep them local because the work is sensitive or confidential, are exactly the sessions this feature skips. So the line does not run between what you do and do not want to share. It runs between where the machine happens to be. Keep your work on your own computer for good reasons and you keep the benefit outside as well.
Who gets it, and who can wipe it
Anthropic gives no numbers at all. No benchmark, nothing about how much Claude retains or how often it gets something wrong. What is fixed is who gets the feature. On for Free, Pro and Max, off for Team and Enterprise until an admin enables it, and still off per user after that.
The least flattering detail sits in the Help Center rather than the announcement. In an organization, an owner can switch memory off for everyone, and that immediately deletes every existing memory entry for every user. Beyond that, memory entries are part of the standard conversation history exports admins can pull under the organization’s retention policy, and that covers incognito chats too. Individual edits to memory are not logged, so an admin sees the export and not the corrections. Organizations on HIPAA, public-sector or custom data retention agreements do not get memory at all.
The caveat
The default shifts from forgetting to remembering. Until now a conversation with a model was accidentally private, simply because nothing was kept. That accident is gone. Memory is on for Free, Pro and Max, which is the large group. You can switch it off, but you have to know it exists, and most people do not read release notes.
The responsibility lands on the user. The sensitive-topics setting is off by default, and Anthropic refuses some categories even when you switch it on: identification numbers, criminal history, immigration status. That is genuine restraint. But guarding the rest is now your job. You have to work through a list of files that grows while you talk, and spot for yourself that something in it is not what you want surfacing in another context. That takes a kind of attention nobody sustains for six months.
What drains away is the re-explaining itself. Explaining is not only overhead. It is the moment I notice that I have not got something straight. When the context is always already there, I skip that moment. I have watched it happen with my own memory files: it goes faster, and I check less often whether what is written there still holds. A memory that updates itself while you talk pushes harder in the same direction.
What I am doing myself
I am switching it on, but not in the same way everywhere.
For the work on this site it stays on. The context there is about tone of voice, about keeping two brands apart, and about how I publish. That is precisely the sort of thing I do not want to explain again, and it already lives in files under version control.
For anything connected to my day job it stays off, and those sessions stay local. That was already the case, and this announcement only sharpens the reason. A shared memory that travels into a cloud session is exactly what you do not want around confidential work.
I am leaving the sensitive-topics setting off. The upside is small and the downside is hard to undo. Anthropic says something useful about it. Switching it on only works forward, and nothing from before that moment is stored retroactively. That is decent design, and it also means waiting costs nothing.
And I will read through the topics list once a month. Not because I expect to find something alarming, but because a list you never open stops being a list within three months and becomes an assumption.
Frequently asked questions
Does this work if I run Cowork locally?
No. The shared memory only works when Cowork runs in the cloud. The Help Center is explicit about it and the blog post does not mention the limit at all. Local Cowork sessions neither read your chat memory nor write anything back to it.
Can my employer see what Claude remembers about me?
Partly. Owners of a Team or Enterprise organization cannot view or edit individual memory entries, but memory entries are included in the standard conversation history exports under the organization’s retention policy. If an owner switches memory off for the organization, every entry for every user is deleted immediately.
Do I need to do anything?
Only if you are on Free, Pro or Max and would rather not have it: memory is on there. On Team and Enterprise it is off until an admin enables it, and still off until you enable it yourself. Sensitive topics are off by default everywhere.
Is this the same as project memory?
No. According to Anthropic each project keeps its own separate memory space and its own summary, apart from other projects and from chats outside a project. The shared memory this announcement is about is the layer around that.
Sources
- Anthropic, “Claude’s memory works everywhere, and you decide what’s in it”, claude.com, 25 August 2026, accessed 25 August 2026.
- Anthropic Help Center, “Use Claude’s chat search and memory to build on previous context”, support.claude.com, accessed 25 August 2026.
- Anthropic Help Center, Release notes, entry for 25 August 2026, support.claude.com, accessed 25 August 2026.
- Amol Avasare on X, announcement post, x.com, 25 August 2026, accessed 25 August 2026.
- Igor Bonifacic, “Claude’s memory now works across both chats and Cowork sessions”, Engadget, 25 August 2026, accessed 25 August 2026.
Checked on 25 August 2026. Every factual claim above comes from Anthropic’s blog post, the Help Center or the release notes; the post on X is the trigger and not the source, since it is personal rather than official. Two things I could not verify myself. I have not been able to test the feature in my own account at the time of writing, so anything about how it feels in daily use is provisional. And the sources contradict each other on one point: the blog post says the memory works everywhere, while the Help Center and the release notes name the limit to Cowork in the cloud. I follow the primary documentation, so: cloud yes, local no.