TRUSTMEM: Learning Trustworthy Memory Consolidation for LLM Agents with Long-Term Memory

2026-06-23

AI Summary

This work addresses the critical challenges in long-term memory updating for large language model (LLM) agents, where updates often omit essential information, corrupt existing memories, or introduce hallucinations, leading to systematic errors. To mitigate these issues, the authors propose TrustMem, a novel framework that introduces, for the first time, a trustworthiness assessment mechanism at the memory transformation level. TrustMem employs a memory transformation verifier to quantify coverage, retention, and faithfulness of updates, constructs preference pairs based on these metrics, and optimizes the memory writing strategy through preference-guided reinforcement learning. The approach significantly enhances memory reliability, achieving state-of-the-art performance on MemoryAgentBench, HaluMem, and Mem-alpha benchmarks. Notably, it improves F1 score by 12.14 points on HaluMem and reduces omissions, corruptions, and hallucinations in memory transformation by 40.1%, 79.1%, and 50.0%, respectively.

Abstract

Large language model (LLM) agents rely on long-term memory to support extended interactions and personalized assistance beyond finite context windows. Existing memory agents actively update external memory through generated write, revise, and delete operations, but these updates may omit important information, corrupt existing memory, or introduce unsupported hallucinated content. Once stored, such errors become persistent system-state failures that can affect future reasoning and generation. In this paper, we propose TrustMem, a framework designed to improve the trustworthiness of memory consolidation. TrustMem relies on a Memory Transition Verifier to evaluate the transition process of memory updates in terms of coverage, preservation, and faithfulness. It further constructs preference pairs among candidate updates under the same memory state, enabling preference-guided reinforcement learning to directly optimize memory updating behaviors. Extensive experiments demonstrate that TrustMem improves both memory utility and reliability: it achieves state-of-the-art results across MemoryAgentBench, HaluMem, and the Mem-alpha validation set, improves HaluMem memory extraction by 12.14 F1 points, and reduces transition-level omission, corruption, and hallucination by 40.1\%, 79.1\%, and 50.0\%, respectively, compared with the strongest baseline for each error type.

Problem

• long-term memory
• memory consolidation
• hallucination
• memory corruption
• trustworthiness

Innovation

• Trustworthy Memory Consolidation
• Memory Transition Verifier
• Preference-Guided Reinforcement Learning
• Long-Term Memory for LLM Agents
• Memory Hallucination Reduction