VerMem: Unified Memory Management for LLM Agents with Local and Global Verifiers

A novel framework named Verifiable Memory (VerMem) has been developed to enhance memory management for large language model (LLM) agents. This framework is outlined in a paper available on arXiv (2608.03137) and introduces a cohesive policy governing long-term memory (LTM), active context, and episodic history through seven fundamental operations: adding, revising, or soft-deleting LTM entries; retrieving LTM into active context; filtering or summarizing active context; and restoring chosen episodic fragments. VerMem undergoes initialization via supervised fine-tuning and employs a three-stage reinforcement-learning curriculum for training. Its primary innovation lies in the implementation of local and global verifiers, which evaluate memory transitions and overall trajectories, respectively. This approach tackles the challenge of inadequate credit assignment seen in current methods that treat LTM and short-term memory (STM) independently. The authors of the paper are not specified in the abstract. The framework aims to enhance long-horizon interactions by allowing agents to retain useful information, manage limited active context, and retrieve past evidence.