Claude Unified Memory: What Changed and How to Use It

Claude's memory upgrade is a big deal, but not in the way the headline suggests. Anthropic now lets Claude Chat and cloud-based Co-Work share the same memory. That means context flows both ways. Tell Claude in Chat that your weekly report starts with decisions and risks, and a Co-Work task can use that structure without you repeating it. What Claude learns while running a Co-Work task also comes back to Chat. The real payoff is continuity: you stop re-briefing the model every time you switch between discussing work and actually doing it.

But there are limits. Claude Code has its own memory system using files like cloud.md and project-scoped automemory. It is not part of this update. Project memories stay separate too. Incognito chats and local Co-Work sessions don't use unified memory. So if your workflow depends on Claude Code or local agents, you won't see the benefit yet.

How memory works now

Claude no longer shows one profile per conversation. Instead, it stores memory in visible topic files you can open, edit, or delete. You'll find them under Settings > Memory on the web or desktop app.

There are two different controls. The first is 'search and reference chats'. That lets Claude look back at old conversations when it needs context. The second is 'generate memory from chats'. That's what creates and updates long-term memory files. You can turn one on without the other. It's worth checking both.

Memory topics cover things like reporting style, recurring projects, or communication preferences. You can click a topic, see what Claude saved, and change it directly. If something is wrong, tell Claude to update it. For example: 'Our weekly report now starts with decisions and risks. Put detailed metrics in the appendix.' The corrected version applies to both Chat and Co-Work. You can delete a topic outright. That beats correcting Claude in every new session.

What to watch out for

There's a separate setting for sensitive topics, and it's off by default. If you turn it on, Claude may remember health conditions, religious beliefs, or similar details. Anthropic says it's not retroactive. It won't scan your old chats and create sensitive memories. Some categories should never be saved, like government ID numbers, financial account numbers, or immigration status.

The practical rule is stricter: don't treat Claude's memory as a password manager or a secrets vault. Use it for working preferences and durable context, not credentials or critical records. That applies to local setups too. When you see the onboarding screen, check the sensitive topics switch instead of accepting defaults.

How to test it

Want to know if the handoff works? Ask Chat and cloud Co-Work the same question, something neither could answer without shared memory. If the answers match, it's working. If not, go back to Settings > Memory and see what was actually saved.

The feature helps most with recurring reports, project status updates, and team communication preferences. If you switch between Chat and Co-Work often, you won't have to repeat the same instructions. Keep an eye on Anthropic's roadmap. The channel's own setup uses CC Switch to access Claude Code from Hong Kong, and whether Anthropic officially supports CC Switch for unified memory is still an open question.