I wanted to understand your design decision to inject context at every turn here: https://github.com/jo-inc/pi-mem?tab=readme-ov-file#context-injection. Maybe injecting once at the beginning is enough and state throughout a session can be handled by the model or a skill?
As i understand it many of the files being injected are likely or at least possible to be updated throughout a session. Unfortunately this will break prompt cache if a mutated result from the conversation history is sent back. This can get expensive and also get subscription users banned. I understand the desire to do it though so i wanted to see if you addressed this in a way that isn't obvious to me in the code to avoid the caching issue.
I wanted to understand your design decision to inject context at every turn here: https://github.com/jo-inc/pi-mem?tab=readme-ov-file#context-injection. Maybe injecting once at the beginning is enough and state throughout a session can be handled by the model or a skill?
As i understand it many of the files being injected are likely or at least possible to be updated throughout a session. Unfortunately this will break prompt cache if a mutated result from the conversation history is sent back. This can get expensive and also get subscription users banned. I understand the desire to do it though so i wanted to see if you addressed this in a way that isn't obvious to me in the code to avoid the caching issue.