Recent advances in large language models have markedly improved agents' ability to model long‑term conversations. Existing approaches largely follow a static heuristic paradigm: information is archived passively without adaptive valuation, so memory management cannot evolve with user needs. To overcome this, we introduce ICML (InteraCtive Memory Learning), a multi‑agent framework that turns memory from a passive store into a learnable, interactive policy.\ \ We first employ a session synthesis pipeline to generate expert data, enabling rapid test‑time adaptation in unseen scenarios. ICML then applies online reinforcement learning where:\
- a Planner agent selectively encodes high‑value information;\
- a Trigger agent dynamically retrieves stored content to optimize response quality. The two agents co‑evolve through continuous interaction feedback.\ \ A delayed reward mechanism propagates future user feedback back to earlier storage decisions, tightly aligning memory policies with user expectations. Experiments show that ICML significantly outperforms strong baselines and uniquely improves response quality as interactions accumulate.\ \ Review: ICML’s combination of multi‑agent collaboration and delayed rewards enables proactive, adaptive memory learning, offering a sustainable improvement path for long‑term conversational systems.