TRUSTMEM framework cuts LLM memory errors by 79% with RL

Large language model (LLM) agents increasingly rely on long-term memory to handle extended interactions and personalized assistance, but current systems suffer from unreliable memory updates. Existing agents can omit important information, corrupt stored data, or hallucinate false content during write, revise, and delete operations. These errors become persistent system-state failures, undermining future reasoning and generation. The new TRUSTMEM framework tackles this by introducing a Memory Transition Verifier that evaluates each memory update across three dimensions: coverage (did it include necessary info?), preservation (did it keep existing correct memories?), and faithfulness (did it avoid adding unsupported content?). It then constructs preference pairs from candidate updates under the same memory state, enabling preference-guided reinforcement learning to directly optimize the agent's memory update behavior.

Extensive testing shows TRUSTMEM delivers state-of-the-art performance across MemoryAgentBench, HaluMem, and the Mem-alpha validation set. Critically, it improves HaluMem memory extraction by 12.14 F1 points and achieves dramatic error reductions: omission drops by 40.1%, corruption by 79.1%, and hallucination by 50.0% compared to the strongest baseline for each error type. This work demonstrates that explicit verification and reinforcement learning can make LLM agent memory more trustworthy, a key step toward reliable autonomous systems that can safely retain and recall information over long periods.