NeuroMem: A Neuroplastic Memory Framework for Lifelong Agents through Delayed Consolidation

With long-term memory, LLM-based agents have demonstrated remarkable capabilities in handling complex long-horizon tasks. However, existing memory frameworks primarily treat long-term memory as a post-admission management problem, emphasizing how experiences should be organized, updated, or controlled. This overlooks a more fundamental stage in the memory lifecycle, i.e., how transient experiences are progressively transformed into stable and task-transferable knowledge. Biological memory consolidation provides a temporal principle for memory formation, whereby long-term memory emerges through the delayed stabilization and selective transformation of recent experience, rather than through its verbatim preservation. Motivated by this, we propose NeuroMem, a neuroplastic memory framework for lifelong agents through delayed consolidation. NeuroMem formulates the formation of task-transferable knowledge in long-term memory as a delayed-reward reinforcement learning problem. Specifically, newly formed memory candidates are first maintained in a transient buffer, where a consolidation policy learns to promote, merge, retain, or discard them only after their delayed utility rewards become observable through subsequent interactions. Experiments on EHRSQL and LoCoMo show NeuroMem achieves stronger performance with a smaller memory set, improving EHRSQL task success by up to 8.7% and LoCoMo F1/BLEU-4 by up to +5.20%/+0.97% over baselines.