Local_ai

I thought running multiple AI agents locally meant loading the model into memory once per agent. Turns out model weights are shared - only the KV-cache scales per request. Here’s what I learned and how consolidating to a single model improved everything.