Hey there! 👋
Welcome back to SavvyMonk, your one-stop for AI and tech news that actually matters. Anthropic merged the memory behind Claude chat and Claude Cowork so context flows both ways, and the fine print underneath that change is more interesting than the headline feature.
Let's get into it.
TODAY'S DEEP DIVE
Anthropic Merges Claude Chat And Cowork Memory With No Option To Separate Them
On 25 August 2026, Anthropic announced that Claude chat and Claude Cowork now run on a single shared memory. Anything Claude picks up while you are talking to it in a chat window is available when you hand Cowork a task, and anything Cowork learns while working comes back the other way.
The company's framing is that you should spend less time re-explaining a project, a manager's preferences, or a client you mentioned last quarter.
The Backstory
Cowork is the agent Anthropic released on 12 January 2026, initially as a research preview for Claude Max subscribers on macOS, built to handle multi-step work with files, browsers and connected applications rather than answering questions in a chat box. It shipped without memory. Chat had memory, Cowork did not, and the two systems never spoke to each other.
That produced the specific annoyance this update targets. You could spend weeks developing a plan in conversation and then have to brief the agent from zero the moment you wanted it to act on any of it.
What Actually Changed Under The Hood
The merge is the visible half. The less advertised half is that Claude now writes memory while you are still typing, saving individual topics as the conversation goes rather than reviewing a chat once it ends and producing a summary. Mention that a deadline moved and the next conversation already knows.
Everything Claude has retained appears as a list of short topic files in Settings, where each one can be opened, edited or deleted, and a correction applied to one topic carries into every conversation after it. That is a meaningful shift in how the system behaves, and it applies to chat regardless of whether anyone ever opens Cowork.
The Cloud-Only Limit
Here is the part the launch messaging does not lead with. Anthropic's own support documentation states that memory is shared between chat and Cowork only when Cowork runs in the cloud, and that Cowork sessions running locally on your computer do not use memory at all.
For anyone who adopted Cowork in its original form as a local agent pointed at folders on their own machine, this feature does not exist. It arrives only when the work is running on Anthropic's infrastructure, which is a fairly significant asterisk on a feature sold as one memory everywhere.
There Is No Split Setting
An Anthropic spokesperson told The Register that there is no option to keep chat memory and Cowork memory apart, because they now use the same underlying system, and that the only way to separate them is to run two different accounts. That matters more than it sounds.
Dario Amodei, co-founder of Anthropic
Plenty of people use chat loosely and personally while treating an agent with file and browser access as something closer to a work tool, and that distinction is no longer available as a setting. The controls Anthropic does offer are a pause, a full reset, and per-topic deletion, all of which apply across both surfaces at once.
What Claude Refuses To Store
By default Claude does not save topics touching health, race, ethnicity, religious beliefs, politics or gender identity. Users who would rather not keep re-explaining those things can switch on an include sensitive topics setting, after which a notice appears above the message box each time something in those categories gets saved. Nothing is captured retroactively, and turning the setting back off removes what was already stored.
A separate category is blocked permanently regardless of preference, covering government identification numbers, criminal history, financial account numbers and immigration status, and Claude tells you when it declines to save something on those grounds.
One Detail Worth Reading Twice
Deleting a conversation does not delete the memory generated from it. Anthropic's documentation is explicit that when a conversation expires or is removed, the related memory entries stay, and the only way to clear them is to delete the individual topics yourself in Settings.
On Team and Enterprise plans the picture inverts, since memory is off until an owner enables it, and switching it back off permanently destroys memory data for every member of the organisation. Memory is also unavailable to organisations operating under HIPAA, public-sector or custom data retention agreements.
Who Gets It
Memory is on by default for Free, Pro and Max accounts across web, desktop and mobile, with iOS and Android users needing the current app build to see any of it. Cowork itself is not available on Free, so the shared-memory half of the announcement starts at Pro.
Claude Code is excluded from the merge entirely, which leaves Anthropic with two of its three main surfaces joined and the developer tool sitting on its own. Anyone still on the legacy memory system has been migrated across, and the option to export the old memory disappears on 9 September 2026.
The Bottom Line
The useful version of this update is real, and handing an agent a task it already understands removes a genuine daily friction. But the way it was announced flattens three caveats that each deserve their own sentence, namely that local Cowork sessions get nothing, that the two memories cannot be separated by anyone unwilling to maintain a second account, and that deleting a chat no longer deletes what was learned from it.
The topic-by-topic settings page is the strongest part of the release, because it is the first time Claude users can read the actual files the system has decided are worth keeping about them. Open it before you decide how you feel about the rest.
AI PROMPT OF THE DAY
Category: Data Audit
"Act as a privacy analyst helping me review what an AI assistant has stored about me. I am going to paste in the list of memory topics it has saved from my conversations. For each one, tell me whether it is work context I would want carried into future sessions, personal information I would rather it forgot, or something ambiguous I should think about. Then flag anything that would be a problem if it surfaced during a task I handed to an autonomous agent with access to my files and browser, and suggest which entries I should delete first."
ONE LAST THING
Every memory feature ships with the same promise, which is that you will repeat yourself less. The quieter trade is that the system needs to keep more, for longer, in one place, and the boundaries you drew informally between different tools stop being boundaries.
Anthropic has handled that better than most by making the stored topics readable and editable rather than leaving them buried. The question worth sitting with is whether you would have written those files the same way yourself.
Hit reply, I read every response.
See you in the next one.
— Vivek
P.S. Know a developer or founder running agents on their own machine? Forward this to them.