Memory manager models are pivotal in agent systems. Existing methods predominantly rely on LLM-judged synthetic question-answer (QA) pairs, making memory valuation dependent on sampled queries and the downstream reader. To address this limitation, we propose CMI-Mem, a reinforcement learning (RL)-based lightweight memory manager model with a hybrid reward that combines downstream QA correctness and intrinsic Conditional Mutual Information (CMI). CMI evaluates the information contributed by new conversational inputs relative to the current memory state without conditioning on a sampled QA query, thereby complementing rather than replacing QA grounding. Our codes are available at: CMI-Mem GitHub, and the CMI-Mem-4B model checkpoint is available at: CMI-Mem-4B Model.
Blogger's Review: CMI-Mem significantly enhances the flexibility and efficiency of memory management by introducing the concept of conditional mutual information, especially in complex conversational scenarios. This innovative approach provides new insights for long-term memory management in agent systems, making it worthy of further exploration and application.