Friday

I created a memory system where both the judge/critic learns to evaluate the demands of the agent task, and the knowledge is CRUD and its saliency is verified. My toy example uses crusoe to dispatch the weak implementer model that creates toy games (nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B), and a stronger judge nvidia/NVIDIA-Nemotron-3-Super-120B-A12B. I also added a vLLM to judge the games and insert its own opinion/taste as a way to mimic esoteric client requests. This is both a shared memory and the N+1.