MemLLM: Finetuning LLMs to Use An Explicit Read-Write Memory