Only solution is to train the issue for the next time.

Architecturally focusing on Episodic memory with feedback system.

This training is retrieved next time when something similar happens

Training is an overkill at this point imo. I have seen agents work quite well with a feedback loop, some tools and prompt optimisation. Are you doing fine-tuning on the models when you say training?