What stood out to me is not that Claude has memory. That part is becoming table stakes. The interesting bit is that Anthropic is now exposing just enough of the machinery for users to tweak it, while still making the whole thing feel oddly scattered across settings, mobile, projects, profile memory, incognito mode, and a few different “don’t remember this” switches.
That’s the tension I keep coming back to with these AI memories: they’re pitched as convenience, but they behave more like a second account layer. If you build with Claude for real work, you don’t just want “remember stuff.” You want to know where that stuff lives, whether it leaks across projects, and what exactly gets retained when you say no. The article suggests Anthropic is moving in that direction, but not all the way there yet.
The sensitive-topics part is the one I’d watch most closely. I’m glad there’s a default off position for health, politics, religion, and the rest, because that’s the only sane default. But the note that older versions or the mobile app won’t show the notice and won’t save the sensitive topic is a bit awkward. Maybe that’s just a transitional rollout issue, but it also reads like one of those implementation details that can quietly create inconsistent behavior across clients. For a feature like memory, inconsistency is the bug that matters.
I also think the project siloing is the most practically useful idea in the whole piece. Memory that bleeds everywhere is usually a liability, not a superpower. Keeping project context separate makes sense for anyone using Claude on multiple clients or multiple jobs. The “add to project” escape hatch for stray chats sounds sensible too, though I’d still want to know how deterministic that is in practice. UI promises are easy; state management is where the mess shows up.
If I were using Claude heavily, I’d probably trust incognito mode for quick throwaway chats, but I wouldn’t treat it as a privacy magic trick. “Not saved, not added to memory, not used to train models” is a clear enough statement, yet that still leaves the normal question of logs, retention, and operational handling, which this piece doesn’t really dig into. That’s not a knock on the article; it just means the word “incognito” may reassure users more than it should.
The whole thing feels like Anthropic trying to make memory legible without scaring people off. Fair enough. But the more memory becomes a product feature, the more I want a plain-language audit trail: what was remembered, where, why, and how to delete it without wondering whether I’ve only edited a summary of a summary.
Reference: How to edit Claude
