NARE: An LLM agent that amortizes reasoning into memory and executable rules (github.com)