Use cheap LLM to summarise memory

[tischi]

We could use a cheap LLM to create a summary of the previous chat to save input tokens for the more expensive actual LLM that we may use for the requests.