Claude Chat and Claude Cowork used to keep separate memory. Tell Claude something in a chat window, and Cowork, Anthropic's agent that works through tasks inside chosen files and folders, had no way of knowing it happened. On August 25, 2026, Anthropic closed that gap. The two products now run on one shared memory system, so a fact learned in one shows up automatically in the other.
That is a narrow fix inside one company's product line, but it says something bigger than a changelog entry. Anthropic just admitted that memory locked inside a single app, even when that app is one of two products it owns and built on the same underlying models, breaks the experience badly enough to justify rebuilding it. The same problem exists at a much larger scale every time someone uses Claude for one task and a different AI tool for the next: none of that context carries over, and no company is racing to fix it. Anthropic solved the small version of the problem. The large version is exactly as broken today as it was on August 24. Most Claude power users are not exclusively Claude users. They run ChatGPT, Gemini, or another assistant for parts of their week too, and every one of those switches still starts from nothing.
Picture how that split played out before the fix. Someone spends fifteen minutes in Claude Chat explaining a client's tone preferences, the format their invoices need to follow, and which vendors to flag automatically. Then they open Cowork to actually process a folder of that client's paperwork, and none of it is there. The agent starts cold, on a task built entirely around context that was just typed out in a different window minutes earlier. That was not a hypothetical edge case. It was the default outcome every time someone used both products for the same piece of work, because the two systems had no shared line of memory to draw from.
What Anthropic Actually Changed on August 25
Before the update, Claude built memory by summarizing a conversation after it ended, then storing that summary for later chats to draw on. Cowork, which launched as a separate, newer product, read from a different pool of context entirely. Someone who spent a week teaching Claude Chat how they like a weekly report formatted still had to explain that same preference to Cowork before it would touch the report. Anthropic's update removes that wall. Chat and Cowork now draw on identical memory, and Anthropic frames the result plainly: Claude "will always remember what it learned in one area, even when you're engaging with it in another."
The shift away from end-of-conversation summaries matters on its own, separate from the Cowork connection. Under the old system, anything said in a conversation that got abandoned, closed early, or left open in a browser tab for days risked never being captured, because the summary step that created memory only ran once a conversation formally ended. Capturing topics as they come up removes that dependency. A preference mentioned in passing, in a chat that never gets a clean close, still gets saved.
In practice, that turns a repeated chore into a one-time setup. Tell Claude Chat once that quarterly reports should open with a one-line summary and end with next steps, then hand Cowork a folder of raw spreadsheets, and the report that comes back can follow that format without the instruction being typed twice. The reverse direction works too. Details Cowork picks up while executing a task, a client's name, a file structure, a stated formatting preference, become available the next time the same topic comes up in ordinary chat.
Why Chat and Cowork Had Separate Memory in the First Place
The split existed because Cowork is a genuinely different, newer product, not a reskin of Claude Chat. Anthropic shipped it as a desktop app in January 2026, positioned as an alternative to Claude Code for people who do not write software: an agent that works inside a folder a user picks, runs multi-step tasks, and keeps working after the laptop closes. It expanded to web and mobile for Max subscribers by July 2026. Anthropic's own usage data, drawn from roughly 1.2 million Cowork sessions, puts business process work like report compilation and spreadsheet reconciliation at about a third of activity, content drafting at roughly 16 percent, and software development, the use case most associated with an Anthropic agent, at under 9 percent. Cowork was built for the parts of a job that are rarely anyone's core responsibility: expense reports, contract review, recurring status updates. That is a different workload from open-ended chat, and it grew its own memory infrastructure to match. Merging it with chat's memory took a deliberate rebuild, not a setting Anthropic simply forgot to flip earlier.
How the Unified Memory Actually Works
- Memory updates during a conversation now, not only after it ends. Anthropic describes the change as adding topics to memory on the fly, as a chat happens, rather than waiting for a closing summary to be generated.
- The flow runs in both directions, but only when Cowork runs in the cloud. Context Claude picks up during a chat becomes available to a cloud Cowork task, and anything that task learns feeds back into ordinary chat; Cowork sessions that stay entirely on a person's own computer do not draw on the shared pool, according to Anthropic's own help documentation.
- Memory is organized by topic rather than stored as raw transcripts. Claude already let people view, edit, or delete individual entries one at a time; that same topic-by-topic control now spans both products instead of covering chat alone.
- There is no setting to keep Chat and Cowork memory separate anymore. The only ways around the shared pool are turning memory off entirely, using an incognito chat, deleting specific entries, or running the two products on separate accounts.
Cowork already worked inside whatever folder or tool a user pointed it at: reconciling a spreadsheet, drafting a batch of proposals, working through a stack of call notes, running with a built-in browser that stays separate from a person's own logins. None of that changes. What changes is the starting context those tasks run on. A finance role that spent months telling Claude Chat how to format a monthly close can now assign that close to Cowork without restating the format. A marketing role that has been refining brand voice notes in chat can hand Cowork a batch of drafts and expect it to already know the voice. The tasks Cowork runs are the same. What feeds into them is no longer limited to whatever gets typed into that one task window.
Think through a single afternoon to see the difference. Someone opens Claude Chat to plan a product launch, listing three deliverables, a launch date, and which teammate owns each piece. An hour later, they assign Cowork the job of drafting the first deliverable from a folder of research notes. Before August 25, Cowork had no idea a launch date or an owner list existed; it just wrote from whatever notes it was pointed at. After the update, Cowork's draft can reference the date and the ownership split without being told again, because that context already lived in memory the moment it was mentioned in chat, not just after the chat window closed.
What Stays Private: Sensitive Topics and Data That's Never Stored
The unification did not loosen what Claude is willing to remember. Health information, race and ethnicity, religious belief, political views, and gender identity stay out of memory by default in both products, the same rule that applied before the merge. Anthropic added a toggle, labeled to include sensitive topics in memory, for anyone who wants Claude to retain that category anyway, and the app shows a notice when it flags something sensitive. A harder line sits underneath that toggle: Social Security numbers, government ID numbers, criminal history, and immigration status are never stored, regardless of what a user enables, along with content that violates Anthropic's usage policies. The merge did not change those rules. It only widened the scope of products they apply to.
Who Gets the Unified Memory, and When
The underlying memory system is on by default across Free, Pro, and Max plans, on web, desktop, and mobile, though iOS and Android users need the current app version for it to take effect. Cowork itself is not part of the free plan; it requires Pro, Max, Team, or Enterprise. On the paid plans that include it, the shared memory specifically applies when Cowork runs in the cloud. Anthropic's own help documentation says plainly that Cowork sessions kept entirely local to a person's computer do not draw on the shared pool. Team and Enterprise accounts need an administrator to turn memory on rather than getting it automatically. Claude Code's memory sits outside the merge entirely, and Anthropic has not said whether that changes. Anyone on a paid plan who already had memory turned on for Claude Chat woke up to it also covering their cloud Cowork sessions; the only decision left to make is what gets kept, not whether the two share a pool in the first place.
Within Claude vs. Across Every Other AI Tool
| Scenario | Within Claude (Chat + Cowork) | Across Different AI Tools |
|---|---|---|
| A deadline changes inside a Claude Chat conversation | Cowork sees the update automatically | Other AI tools never see it happen |
| Cowork finishes a task using files in a folder | Chat can reference what Cowork learned | Every other tool starts from zero |
| Someone wants Chat and Cowork memory kept apart | Not possible without a separate account | Not applicable, memory never crossed over |
| Someone switches from Claude to a competing AI tool | Not applicable, memory stays inside Anthropic's apps | None of that Claude memory transfers |
| A workweek runs through three or four different AI tools | Not applicable, unified only within Claude | Each tool starts the week knowing nothing |
The Gap This Update Doesn't Close
Read the announcement closely and the boundary is obvious. Everything Anthropic fixed happens inside Anthropic's own two products. A fact told to Claude Chat now reaches Claude Cowork. It still does not reach ChatGPT, Gemini, Perplexity, or whichever tool gets opened for the next task. Most people do not run their entire workday through a single AI product. They pick whichever tool fits the task in front of them, a habit that makes the real memory problem, re-explaining who someone is and what they are working on every time a different tool opens, the default experience for almost everyone, Claude users included.
Consider a common week. Research happens in one AI tool because it cites sources well, drafting happens in Claude because the tone comes out right, and a spreadsheet task gets handed to Cowork because it can work unattended in the background. Each leg of that chain is a reasonable tool choice on its own. None of the tools know what the others just did, so every handoff starts with a recap that has nothing to do with the actual work.
There is a broader signal worth sitting with too. Unifying memory across two products built by the same company, on related models, with one engineering organization able to touch both codebases, still took Anthropic a dedicated rebuild rather than a quick fix. Getting two competing companies with no shared incentive to cooperate to make memory portable between their separate products is a harder version of a problem one company just found difficult to solve for itself.
None of that is a criticism specific to Anthropic. It is a structural fact about how AI products get built. Each vendor keeps its own account system, its own memory format, and its own reason to hold a person's history inside its own app rather than making it portable. A company that made switching away from its product easier would be acting against its own retention numbers, and that incentive does not soften just because two products share a parent company. Whether any vendor eventually opens that door on purpose, or whether it takes outside pressure to build it, is a separate question from whether the underlying problem is real. Anthropic's own move this week suggests it is.
This is the specific gap an external memory layer like MemX (memx.app) is built to close. Instead of living inside one vendor's app, MemX sits outside any single AI product and holds the context a person builds up over time, so a preference or project detail saved while using one AI tool is available the next time a different one is opened. MemX describes its approach to privacy as private by architecture: per-user isolation, customer-managed encryption keys, encryption at rest, and processing on-device where possible, with no training on user data. That trade-off is worth naming plainly: an external layer means one more account and one more thing to trust, in exchange for context that does not evaporate every time a different tab opens. It will not matter much to someone who lives entirely inside Claude Chat and Cowork, Anthropic just solved that case directly. It matters for the far more common case: a workday that runs through three or four different AI tools that still do not talk to each other. The two approaches are not mutually exclusive. Someone can leave Anthropic's native memory switched on for the work that happens entirely inside Claude, and still route the rest of a multi-tool week through something built to survive a tab switch.
Frequently Asked Questions
01Does Claude Cowork remember what I said in Claude Chat?
Yes, but only on a paid plan and only when Cowork runs in the cloud. Cowork itself is not available on the Free plan, only Pro, Max, Team, and Enterprise. On those plans, a fact mentioned in Claude Chat since the August 25, 2026 update is available the next time a cloud Cowork task runs, and vice versa.
02Can I keep Claude Chat and Cowork memory separate?
Not through a setting. Anthropic removed that option when it merged the two systems. The available workarounds are turning memory off entirely, using incognito chats, deleting specific memory entries, or using two separate accounts, one for each product.
03What information does Claude never store in memory?
Social Security numbers, government ID numbers, financial account numbers, criminal history, and immigration status are excluded permanently, even for users who opt in to sensitive topics. Health information, ethnicity, religion, political views, and gender identity stay out of memory by default unless specifically enabled.
04Does the new Claude memory update work with ChatGPT or Gemini too?
No. The unification announced on August 25, 2026 only connects Claude Chat and Claude Cowork, both Anthropic products. Memory does not carry over to other vendors' tools, which is the specific cross-tool gap that external memory layers like MemX are built to address instead.
05Is the unified Claude memory feature on by default?
The underlying memory system is on by default for Free, Pro, and Max plans on web, desktop, and mobile. Cowork itself needs a paid plan (Pro, Max, Team, or Enterprise), and shared memory only applies when Cowork runs in the cloud. Team and Enterprise need an administrator to enable it; Claude Code's memory sits outside this merge.
