TRUSTMEM: Learning Trustworthy Memory Consolidation for LLM Agents with Long-Term Memory
06:00 · June 25, 2026 · arXiv cs.AI RSS
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.
Summary
Large language model agents increasingly depend on external long-term memory to maintain continuity across extended interactions that exceed fixed context windows. Current approaches allow agents to actively edit this memory through write, revise, and delete operations, yet these updates frequently omit key details, alter existing entries incorrectly, or insert unsupported content that persists as latent errors in future reasoning.
TrustMem addresses these risks by shifting evaluation from final task outcomes to the individual memory transitions themselves. A dedicated Memory Transition Verifier assesses each candidate update against three criteria: coverage of the incoming information, preservation of prior correct entries, and faithfulness to the source material without hallucination. The framework then constructs preference pairs among multiple possible updates under identical memory states and applies preference-guided reinforcement learning, specifically a variant called Transition-Ranked GRPO, to refine the policy that generates the edits.
Experiments across MemoryAgentBench, HaluMem, and the Mem-α validation set show that the resulting policy improves both downstream utility and operational reliability. TrustMem reaches state-of-the-art scores on all three benchmarks, raises memory-extraction F1 by 12.14 points on HaluMem, and lowers transition-level omission, corruption, and hallucination rates by 40.1 percent, 79.1 percent, and 50.0 percent respectively relative to the strongest prior baselines for each error type.
Why it matters
This research is highly relevant for Dutch AI researchers and developers focusing on trustworthy and reliable AI systems, a key priority in the Netherlands and the EU. Improving LLM memory consolidation directly addresses enterprise needs for robust, hallucination-free AI agents capable of long-term interactions.
