> how they think they can force an LLM to pull from the memory bank instead of the training data

You have to implement procedurality first (e.g. counting, after proper instancing of ideas).