Causal-Salience Memory: Causality-Aware Memory and Explainable Preference Alignment for Personalized Agents

Personalized agents powered by large language models (LLMs) typically rely on recency-based or embedding-similarity retrieval to access stored user memories, but such approaches conflate correlation with causation, leading to unstable recommendations and opaque decision rationales. We propose Causal-Salience Memory (CSM), a memory framework that augments each memory entry with explicit causal schema—recording cause, trigger condition, and effect—alongside a dynamic salience score. CSM further introduces a causal consistency checker that detects conflicts between newly generated suggestions and established causal knowledge before responses are delivered. We evaluate CSM on two public benchmarks—the Multi-Session Chat (MSC) corpus and the PersonalDialog dataset—measuring preference hit rate, recommendation consistency, and explanation quality. Experimental results show that CSM achieves a preference hit rate of 84.2% at session~10 (compared to 68.5% for embedding-only retrieval), improves long-term consistency by 35.8% relative, and produces explanations rated 4.23/5 by GPT-4 evaluation. The design adds 118∼ms of per-query overhead. These results suggest that structured causal memory offers a practical path toward more stable and interpretable personalized agents.