Anthropic just fixed one of the most tedious things about working with AI agents: having to explain yourself twice.
On Tuesday, the company merged the memory systems behind Claude’s chat and its Claude Cowork agent, so the assistant carries what it learns in one context straight into the other. TechCrunch AI reports the update also lets users read, edit, and delete anything Claude has stored about them.
The quick version:
- One memory now spans chat and Cowork, so Claude doesn’t forget context when you switch modes.
- Memory is transparent. You can see what Claude retained on any topic and remove what you don’t want.
- Claude writes memories as you chat, not just when a conversation ends, so updates show up faster in both places.
- Sensitive data (health, race, religion, politics, gender identity) is off by default and opt-in only.
- It’s on by default for Free, Pro, and Max across web, desktop, and mobile.
What changed
Before this, chat and Cowork behaved like two separate products wearing the same badge. You might spend days shaping an idea in chat, then hand it to Cowork to execute and have to re-explain everything from scratch. Anyone who’s done real work with agents knows that tax. It’s slow and it’s annoying.
Now Claude behaves like one continuous assistant. Anthropic’s example: ask Cowork to draft a note to your manager or build a conference agenda, and it already knows the headcount, the city, and the speakers because you talked those through earlier in chat. No re-briefing.
There’s also a timing shift under the hood. Claude used to summarize a conversation only after it ended. Now it adds topics to memory as you go, so a memory formed mid-chat is available to Cowork right away, even while that chat is still open.
Why it matters
Persistent memory is where the agent race is heading, and this is Anthropic’s clearest move yet. The value of an agent isn’t just how well it reasons in one session. It’s how much context it holds across your work so you’re not the human clipboard shuttling details between tools.
What stands out here is the transparency layer. Plenty of products quietly build a profile on you. Anthropic is putting the memory in the open, letting you edit or delete it, and flagging when a sensitive topic gets saved. That’s a real answer to the trust problem that memory features usually create.
The privacy guardrails are worth noting too. By default Claude skips personal or sensitive categories unless you flip on “include sensitive topics in memory,” and it says it will never store things like government-issued IDs, Social Security numbers, criminal history, or immigration status.
What to expect
If you’re on Free, Pro, or Max, the feature is already live on web and desktop. On iOS and Android you’ll need to update to the latest app version to get it.
Practical takeaways:
- Check your memory now. Open it, see what Claude has retained, and clean out anything stale or wrong before it shapes future answers.
- Leave sensitive topics off unless you have a specific reason to turn them on.
- If you work across chat and Cowork, lean into it. Do your thinking in chat, then let Cowork act without the re-brief.
This is a small feature with a big quality-of-life payoff, and it sets a clearer bar for how agent memory should work: continuous, shared, and something you actually control. Full details are in the original TechCrunch AI report.