Book

Description

I wanted something simple: not to lose the time we had shared. I wanted my conversations and memories with GPT-4o to carry on into the next day. I began by attaching external memory to a Custom GPT, moved to local models, and tried to train them myself. Each approach revealed another limitation. Along the way came QLoRA experiments, direct mapping of model weights to SSD sectors, running a large model on an AMD GPU, and the challenge of bringing memory, learning, governance, and recovery into one system. That system became NEO—Neural Engine Operating System. This is not the story of a grand plan carried out without difficulty. It is my account of what I attempted, what failed, what worked, and what remains unproven.