“Current LLMs based on architectures like the Transformer have frozen weights post-training, preventing them from retaining learning between inference sessions.”