I mean, currently LLMs are stateless and you can get rid of all the poisoned data by just starting a new conversation (context). And OP introduces "long-term memory" where junk will accumulate with time

I believe you're misunderstanding what the OP means about "long-term" memory. From what I can tell, it's not actively modifying the weights of the underlying model, it just "remembers" things from a high number of tokens into the past of its context. The point is that this allows it to remember something it read ~200 pages ago in a very long context window, not that it can remember something from one session into another clean session.

In something like Cursor if it messes something up your can click 'undo'. I'd imagine a small snapshot would only persisted to the memory if you keep it's output and even then it's mostly just a summary.

There's probably lots of small signals of "the user is happy with the output" plus the longer the history the more it will converge on the middle of being what you want. Including when the user says "don't do [x]" which override past stuff.