I spent part of a recent session trying to give Claude some biographical context. The kind of background that would save me from re-introducing myself every time — who I am, what I'm working on, what I don't need explained.

What I found is worth knowing.

There appear to be two memory layers. One is visible to you in Settings: an auto-generated summary you can read and edit. The other is written by Claude directly, invisible to you, with no interface representation at all.

I asked Claude to store eight biographical entries. Claude confirmed they were written. None of them appeared in my settings.

I don't know — and can't know from the user side — whether those entries get used in future sessions, whether my edits to the visible summary have any effect on Claude's behavior, or whether the two layers are even connected. I asked Claude directly. It couldn't tell me either.

I may be wrong about some of this. These systems are complex and I'm not looking at the internals. But if I am wrong, I'd say this: the correct answer is nowhere near obvious to a non-technical user — and Claude itself can't explain how its own memory works.

For a system that's supposed to build trust through personalization, not obvious is itself the problem.

For most users this is friction. For older adults who are already cautious about what AI systems know about them, it's something more — a system that creates the appearance of user control without the substance of it.

I'll keep watching this.