Long-Range Tasks Using Short-Context LLMs: Incremental Reasoning With Structured Memories