Researchers have developed CMI-Mem, a lightweight memory manager for agent systems that uses reinforcement learning to improve long-term memory management. Unlike existing methods, CMI-Mem evaluates the information contributed by new inputs relative to the current memory state without relying on sampled question-answer pairs. This approach provides per-operation supervision for relevant, non-redundant memory construction and has been shown to improve transfer across different scenarios and reduce training and inference time. --- Why it matters: This matters because it addresses a critical challenge in AI: managing long-term memory efficiently. CMI-Mem's ability to evaluate new information relative to the current state can lead to more accurate and robust agent systems, especially in applications where memory management is crucial. Source: https://arxiv.org/abs/2607.20553