Two papers accepted to ACL 2026: Memory management for LLM agents and Proactive reasoning in large language models