LLM-State: Expandable State Representation for Long-horizon Task Planning in the Open World