Infinite Retrieval: Attention enhanced LLMs in long-context processing