Memory-Augmented LLM Agents for Long-Term Personalized Decision Making

Abstract

Large language model agents are increasingly proposed as intermediaries for complex, longitudinal decision support, yet their default reliance on finite context windows and stateless inference limits their capacity to maintain coherent personal histories. This paper examines memory-augmented LLM agents as socio-technical infrastructures for long-term personalized decision making. It argues that memory is not a peripheral feature but a foundational layer that shapes privacy, accountability, robustness, fairness, and institutional legitimacy. The analysis distinguishes episodic, semantic, procedural, and affective memory functions and considers how retrieval, summarization, consolidation, and forgetting mechanisms interact with model inference, tool use, and human oversight. Structural trade-offs are analyzed across latency, cost, explainability, consent, data minimization, and failure containment. The paper further discusses nonstationarity and personal drift, showing that long-term personalization requires mechanisms for revision, contestation, and temporal validity rather than simple accumulation. Governance challenges include auditability, liability allocation, cross-border data flows, and the risk that opaque memory stores become unaccountable shadow records. Deployment models are compared across cloud, edge, and hybrid architectures, with attention to healthcare, finance, education, and public services. The paper concludes that sustainable memory-augmented agents require layered architectures, explicit memory contracts, continuous evaluation, and policy frameworks that treat personalization as a governed relationship rather than an engineering optimization. Such systems can support better decisions only if they remain contestable, resource-aware, and aligned with the long-term interests of the people they serve.