ICML Multi-Agent Framework for Interactive Memory Learning

The ICML framework replaces static memory archives with a learnable policy using a Planner agent for selective encoding and a Trigger agent for dynamic retrieval. It utilizes online reinforcement learning and a session synthesis pipeline to enable test-time adaptation for long-term conversational agents.