I'm not sure about Python, but in TypeScript streaming is just a for loop. Similarly, if you're using OpenAI, actions are already built into their client librar

I'm not sure about Python, but in TypeScript streaming is just a for loop. Similarly, if you're using OpenAI, actions are already built into their client library and are trained against in the models, so there's nothing special you need to do. Memory is just an array! They have some neat extras like sample selectors that can be useful — although even then, if you have so many examples you need a sample selector, finetuning gpt-3.5 is often better than using a sample selector with gpt-4 (and is considerably cheaper) in my experience.