Verifiable Memory: Learning Unified Memory Management with Local and Global Verifiers for Large Language Model Agents

2026-08-04

AI Summary

This work addresses the challenge that large language model agents struggle to jointly manage long-term and short-term memory during extended interactions, compounded by the lack of fine-grained credit assignment for memory operations in existing approaches. To this end, we propose VerMem, a framework that models long-term memory, active context, and episodic history as distinct states and governs seven atomic memory operations through a unified policy. VerMem introduces, for the first time, a hierarchical credit assignment mechanism featuring cooperative local and global verifiers: the local verifier assesses the feasibility of memory transfers, while the global verifier ensures evidential consistency and final memory quality. Combined with supervised fine-tuning initialization and a three-stage reinforcement learning curriculum, VerMem significantly outperforms strong baselines across five benchmarks and two backbone models, achieving the best efficiency–performance trade-off in token-constrained interactive settings.

Abstract

Large language model (LLM) agents must retain reusable information, control a bounded active context, and recover earlier evidence during long-horizon interaction. Existing methods commonly optimize long-term memory (LTM) and short-term memory (STM) separately, while unified policies are often trained primarily with trajectory-level feedback, which provides weak credit for individual memory decisions. We present Verifiable Memory (VerMem), a framework that represents LTM, active context, and episodic history as distinct states and controls them with one memory operation policy. Seven atomic operations let the policy add, revise, or soft-delete LTM entries; retrieve LTM into the active context; filter or summarize the active context; and restore selected episodic fragments. VerMem is initialized by supervised fine-tuning and trained with a three-stage reinforcement-learning curriculum. The local verifier scores executable memory transitions, and a global verifier assesses evidence coherence and terminal-memory consistency after task completion. These scores are combined with programmatically computed task, evidence-recall, efficiency, and constraint signals through hierarchical credit assignment. The verifiers are used only during training. Across five benchmarks and two LLM backbones, VerMem achieves the best result on the vast majority of reported metrics and consistently outperforms strong memory baselines. Under controlled online-token budgets on three interactive benchmarks, it also achieves the strongest efficiency--performance frontier among the compared methods. Code is available at https://github.com/Sun-SYSU-24/VerMem.

Problem

• Large Language Model Agents
• Memory Management
• Long-term Memory
• Short-term Memory
• Verifiable Memory

Innovation

• Verifiable Memory
• unified memory management
• local and global verifiers
• atomic memory operations
• hierarchical credit assignment