VerMem Unifies LLM Agent Memory Management with Verifiers.

This paper introduces Verifiable Memory (VerMem), a novel framework designed to enhance the memory management capabilities of Large Language Model (LLM) agents. Existing approaches often treat long-term memory (LTM) and short-term memory (STM) separately, and unified policies typically rely on weak, trajectory-level feedback for individual memory decisions. VerMem addresses these limitations by representing LTM, active context, and episodic history as distinct states, controlled by a single, unified memory operation policy. The framework employs seven atomic operations, allowing the policy to add, revise, soft-delete LTM entries, retrieve LTM into active context, filter/summarize active context, and restore episodic fragments. VerMem is trained using a three-stage reinforcement learning curriculum, incorporating both local and global verifiers. The local verifier scores executable memory transitions, while the global verifier assesses evidence coherence and terminal-memory consistency, combining these with other signals for hierarchical credit assignment. Crucially, these verifiers are only used during training. Across five benchmarks and two LLM backbones, VerMem consistently outperforms strong memory baselines, achieving superior results on most metrics. It also demonstrates a stronger efficiency-performance frontier under controlled online-token budgets, making it a highly effective solution for long-horizon interactions.