Anthropic on Tuesday merged the memory systems behind Claude chat and Claude Cowork, so that anything the model learns in one surface carries over to the other. The change ends the most common daily friction for Cowork users: having to rebrief the agent on projects, people, and details that were already discussed in chat. Alongside the merge, Anthropic is exposing the memory itself — users can now read, edit, or delete what Claude has stored on any topic.
The feature is on by default across Free, Pro, and Max plans on web, desktop, and mobile. iOS and Android users need to update to the latest app version to see it. There is no separate pricing tier for the unified memory; it ships to every paying and free user simultaneously.
Before this update, Claude effectively behaved as two products under one brand. A user might spend a week refining a project idea in chat — the headcount for a conference, the speaker list, the host city — and then hit Cowork to draft the actual agenda, only to find Cowork knew none of it. Anthropic's example in the announcement is exactly that scenario: a manager update or conference agenda that Cowork can now draft with the context already loaded.
Key facts
- 01Anthropic announced Tuesday that Claude chat and Claude Cowork now share a single memory system across web, desktop, and mobile.
- 02Users can read, edit, or delete anything Claude has stored, on any topic, through a new memory inspector.
- 03Memory is on by default for Free, Pro, and Max plans; iOS and Android users need the latest app version.
- 04Sensitive categories — health, race, religion, politics, gender identity — are excluded unless users toggle on 'include sensitive topics in memory.'
- 05Claude will never store government IDs, Social Security numbers, criminal history, or immigration status.
The mechanics changed too. Claude previously wrote memories by summarizing a conversation at the end of it. Under the new system, Claude adds topics to memory as the conversation happens, which means the chat side and the Cowork side stay in sync in near real time — even mid-conversation. That matters for agentic work where a user is bouncing between planning in chat and executing in Cowork within the same session.
The memory inspector is the more consequential piece for enterprise buyers. Users get a direct view of what the model has retained and a control to remove it. That addresses a persistent complaint about AI memory features across the industry: they behave as a black box, and users have no reliable way to audit what an assistant thinks it knows about them.
Anthropic set the sensitive-data defaults conservatively. Claude will not, by default, store health information, race, ethnicity, religious beliefs, political views, or gender identity. Users who want the model to retain that context — for legitimate reasons like health tracking or accessibility — can toggle on "include sensitive topics in memory," and Claude will notify the user each time a sensitive topic is saved.
Some categories are never stored regardless of user preference: government-issued ID numbers, Social Security numbers, criminal history, immigration status, and anything else that would violate Anthropic's acceptable use policy. That's a harder line than most consumer AI products draw, and it follows Anthropic's broader pattern of building policy constraints into the product surface rather than relying on prompt-level guardrails alone.
The competitive frame here is OpenAI, which shipped persistent memory in ChatGPT well before Anthropic and has iterated on it repeatedly. Anthropic's late-mover version is more transparent by design — the read-edit-delete interface and the sensitive-topic notifications go further than what ChatGPT surfaces to end users. That trade-off, shipping later with tighter user controls, is consistent with how Anthropic has positioned Claude against OpenAI on the enterprise side.
The open question is how well memory quality holds up at scale as users accumulate months of context across both products. Memory systems degrade when the retrieval layer surfaces stale or contradictory facts, and unifying two previously separate stores multiplies the chances of that happening early. Anthropic has not shared retrieval accuracy numbers or how it resolves conflicts between chat-side and Cowork-side memories written about the same entity.
For Anthropic, unified memory turns Claude from two adjacent products into a single assistant with continuity — which is the actual prerequisite for the agentic use cases Cowork is built for. Long-horizon agent work only pays off when the agent remembers the prior steps. The read-edit-delete controls are also a quiet competitive lever: the more auditable the memory, the easier the enterprise procurement conversation gets. That's the market Anthropic is optimizing for, and this release fits it.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
