AdaMem: Learning What to Remember for Personalized Long-Horizon LLM Agents

AdaMem introduces a novel approach for managing long-term memory in Large Language Model (LLM) agents by employing a role-specific Memory Policy that learns what to retain based on user feedback. The method enhances question-answering accuracy by up to 9% while reducing memory size by 9% compared to traditional uniform memory systems. This is significant for practitioners as it addresses the challenges of memory bloat in personalized interactions, optimizing both performance and resource usage in LLM applications.