Claude's new memory: what changed, and what it still can't do

The three changes

• One memory across chat and Cowork. The desktop agent that builds your documents now draws on the same recollection as the chat window where you explained how your team works. Previously those were separate rooms.
• Editable topics. Everything Claude remembers is listed under Settings → Memory → Topics. You can read each entry and delete or change it. No more guessing what it thinks it knows.
• Sensitive subjects off by default. Health, beliefs, race, ethnicity, gender identity and political views stay out of memory unless you switch them on. On by default for Free, Pro and Max; off by default for Team and Enterprise.

This follows the July 10, 2026 change that replaced Claude's daily memory summary with individual categorised entries. August made those entries yours to see.

Why it matters more than it sounds

Most memory complaints are not "it forgot" — they are "it remembered the wrong thing and I couldn't reach in and fix it." A one-off remark becomes a standing belief; a project you abandoned in March keeps shaping answers in August. A visible, editable list is the difference between a system you can correct and a system you can only work around.

The sensitive-topics default is a real piece of product ethics too. Most memory systems treat every sentence as equally storable. Deciding that a health worry mentioned once should not silently become a permanent fact about you is the right call, and it should become the industry norm.

The gap

Everything above happens inside Anthropic. The topics list is a summary Anthropic wrote about you, stored on Anthropic's servers, kept under Anthropic's retention rules, readable only by Anthropic's models. You can edit it. You cannot take it.

That distinction stops being abstract the first time you switch tools, hit a deprecation, lose an account, or simply want the same context in two places. Nothing in the August update makes what Claude knows about you readable by ChatGPT or Gemini — see how to move AI memory between providers for what that actually takes today.

There is a second, quieter gap: provenance. A topics list tells you what Claude believes. It doesn't tell you why — which message it came from, whether you said it or the assistant inferred it, whether it was a plan or an accomplishment. We wrote about that in did you actually say that?

What to do this week

• Open Settings → Memory → Topics and actually read the list. Most people find at least one entry that is stale or wrong.
• Decide deliberately about sensitive topics rather than leaving the default unexamined in either direction.
• Keep your own copy of the source material — the conversations and documents the summary was drawn from. That is the part no provider gives back.

The honest read

Every provider adding memory is good news for the idea that an AI should know you. It also sharpens the only question none of them answer: whose copy is it? Anthropic just made their copy far better. It is still their copy.