RICO-Mem: Conflict-Aware Memory Management for Long-Term LLM Agents

Abstract

Long-horizon memory is essential for Large Language Model (LLM) agents, yet existing memory systems still face four key challenges: context truncation breaks links between related memories, LLM-based memory summarization introduces hallucinations, concept drift is often mistaken for factual contradiction, and manually designed rules fail to generalize to unseen memory conflicts. To address these limitations, we propose RICO-Mem, a rule-guided memory architecture built upon a temporal memory graph. RICO-Mem introduces a temporal predicate that distinguishes concept drift from factual contradictions through lifecycle-aware reasoning. It further incorporates a cross-graph ϵ-edges validator that verifies structural consistency with zero additional LLM token overhead. To preserve essential reasoning paths during memory compression, we design a constraint-locked Safe-GC strategy, together with a RICO Learner that automatically discovers high-quality rules through counterfactual impact analysis. Extensive experiments on five public benchmarks demonstrate the effectiveness of RICO-Mem. It achieves the highest answer accuracy on all five datasets, including 0.896 on ECON, outperforming the strongest baseline by 3.7%, while reducing LLM token usage by up to 97%. RICO-Mem also attains a contradiction detection recall of 92.5–100%, substantially higher than baseline methods (54.5–81.8%). Moreover, the proposed ϵ-edges validator reaches a faithfulness score of 0.967 without any validation tokens, whereas LLM self-correction requires 14,560 validation tokens yet achieves only 0.784 faithfulness.