Support interceptor pattern for LLM responses

[elroy-bot]

I'd like to implement a memory extension for llm. The best fit I see for this is adding a custom "model", which queries memories and appends to the conversation before passing it all along to another model.

The ux for selecting the underlying model would be not very good though - the user would need to specify a "model" which is really a model wrapper, and I'd have to re-implement functionality to select a model

Would would be easier/cleaner is to just have an execute function that basically just calls super().execute() with the same params it received