Enhancing Long Context Performance in LLMs Through Inner Loop Query Mechanism