DOMR-Mem: Detecting, Repairing, and Preserving Long-Term LLM Memories

Abstract

Large language model (LLM) agents rely on external memory to support long-horizon tasks, yet existing memory systems struggle to distinguish beneficial from harmful information. Similarity-based retrieval introduces conflicting memories into the context, while memory cleanup and summarization often discard rare but reasoning-critical knowledge or break logical dependencies. We propose DOMR-Mem, a memory architecture that extends relational database outlier management rules (OMRs) to natural-language memory. DOMR-Mem defines predicate-driven classification rules that partition memory into good, bad, and ugly categories, enabling deterministic memory management without relying on LLM-based conflict resolution. The architecture integrates task-driven context pruning to remove harmful memories during retrieval, generative memory repair to maintain long-term consistency, and outlier-aware consolidation to preserve rare but important knowledge during memory compression. We implement DOMR-Mem on GPT-4o-mini and evaluate it on six benchmarks (ECON, ContraDoc, RTE, LoCoMo, LongMemEval, and HaluMem). Experimental results demonstrate state-of-the-art performance in contradiction detection and long-term memory question answering, achieving 71.88% causal reasoning accuracy, 3.86 context perplexity, and strong scalability for automatic rule discovery.