One fascinating challenge with local LLM-powered NPCs is context management—balancing coherent, memory-aware conversations without overloading limited hardware. I’d love to know how you approached this in your Godot 4.x project.
Thanks for asking! Currently I keep the full conversation history, but I'm planning to implement a rolling history limited to the last 20 messages to better manage memory on resource-constrained devices.
My main context management right now is through dynamic system prompts that include current learning progress and subject focus, which keeps the AI responses relevant and educational.