Investigating Context-Faithfulness in Large Language Models: The Roles of Memory Strength and Evidence Style