2. MEM1 This work introduces MEM1, an RL framework for training language agents that operate efficiently over long-horizon, multi-turn tasks by learning to consolidate memory and reasoning into a compact internal state.

→ View original post on X — @dair_ai