I notice that the use of tokens on plugins and therefore memory works differently and not as effectively as native gpt4. For example if I call my plugin and receive results back. Next ask for something that you already know if in the returned json and most times it will call api again. This is a huge waste of resources. For some reason it does not seem to ‘remember’ the contents of the most recent json pulled. However if you tell it in the prompt to give you something from the previous json - it will. For inexperienced users most will not know to do this and I would suggest it should work this way by default. Are others finding a similar pattern and disconnect? (in case you wonder I have tried this with different sizes of returned json)
I’m not sure what you’re driving at, GPT-4 has no memory.
It doesn’t know what is in context unless it processes the context, so it is more token efficient to call the API after processing the prompt.