A correction that doesn't stick feels like it can't learn — but it's almost always about where the correction lives, not whether it was understood.
A correction made mid-conversation applies to that session — it doesn't automatically carry forward to a new one, because a fresh session doesn't have that conversation's memory.
"That's not quite right" is a weaker signal than stating the actual rule. Without the specific rule spelled out, the same category of mistake can resurface in a slightly different form.
Two bugs can produce identical symptoms with different root causes. What looks like "the same mistake again" is sometimes a genuinely new issue that happens to look familiar.
The free guide covers the permission model and a first real workflow — good CLAUDE.md habits from day one.
Get the Free Guide →Not automatically — a correction made in conversation applies to that session. If it should apply every time, it needs to live in CLAUDE.md, not just be said once.
No — it's a sign the feedback needs to live somewhere persistent (CLAUDE.md) rather than only in one conversation's memory, which naturally doesn't carry forward to a new session.
That works but doesn't scale — write it into CLAUDE.md once instead, so every future session starts already knowing it.