Embedding Moral Experience: A Neuro-Symbolic Agent with Structured Ethical Memory for Internally Justifiable Decision-Making

Abstract

Empowering artificial intelligence to make morally sound decisions is crucial, as these systems are increasingly deployed in ethically delicate fields like autonomous driving and healthcare. Achieving robust moral decision-making remains a major research challenge, as current methods vary from entirely data-driven models to rigidly rule-based systems. These approaches share intrinsic drawbacks regarding internal consistency, explainability, and ethical justification. In this paper, we introduce MORAL-MEM, a neuro-symbolic prototype model for moral reasoning that integrates structured ethical memory. Unlike existing approaches that treat memory as transient or rely solely on statistical correlation, MORAL-MEM is technically distinct in its use of a structured quadruple representation (Situation, Action, Affective Valence, and Lesson). This architecture enables the agent to not only retrieve past cases but to symbolically verify the logical coherence between the chosen action, the abstracted moral lesson, and the simulated affective feedback, thereby generating internally justifiable natural-language explanations. We evaluate MORAL-MEM using 20 cases from the Moral Machine Dataset against two baseline models: a rule-based decision system and a neural decision system. Performance is measured using the proposed Ethical Coherence Score (ECS), an objective, rule-based metric that evaluates the coherence between the chosen action and the deduced lesson, the coherence between the action and the conveyed affective valence, and the existence of an internal justification. Experimental results show that while both baseline models score 0.00 reflecting their lack of clear reasoning and justification methods rather than necessarily poor decision quality MORAL-MEM achieves a perfect ECS of 1.00 across all assessed tasks, demonstrating its superior capacity for explainable moral reasoning.