It is not remembering wrong. It never knew.
When you reject an idea, the rejection usually happens in conversation and the thing you rejected never becomes a file. So there is no document saying "we considered this and decided against it", and the next model — or the same one in a new session — reads the same codebase, reaches the same conclusion, and proposes it again with complete confidence.
The more reasonable the idea was, the more often this happens. Bad suggestions get made once. Good-but-rejected ones come back every few weeks.
Memory features store what was said. A rejection is not a fact about your project that can sit alongside the others — it is a decision *about* a fact, and storing it as though it were true is worse than forgetting it.
So most systems discard it. Which means the same wrong idea is genuinely new information every time it comes up.
Canonly keeps refusals as first-class records with the reasoning attached, and answers the next proposal of substantially the same thing with the decision, its date and why — across different AI tools, in different words, months later.