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

Short summary

TrustMem is a framework for improving the trustworthiness of long-term memory in LLM agents by verifying memory updates for coverage, preservation, and faithfulness. Using preference-guided reinforcement learning, it achieves state-of-the-art results and reduces hallucination errors by 50%, omission by 40.1%, and corruption by 79.1% compared to baseline approaches.

• •Addresses persistent memory corruption, omission, and hallucination in LLM agents using Memory Transition Verifier
• •Reduces hallucination (50%), omission (40.1%), and corruption (79.1%) errors vs baselines
• •Uses preference-guided reinforcement learning to optimize memory consolidation behavior

Generated with AI, which can make mistakes.