Tuning LLMs by RAG Principles: Towards LLM-native Memory