Limits of Applicability of Artificial Intelligence in Public Administration: Memory of LLM-Based Agents vs. Human Memory

In recent years, large language models and agent systems based on them have been considered a promising tool for the digital transformation of public administration. However, the practical effectiveness of such systems is determined not only by the quality of text generation, including grammatical correctness, coherence, and general awareness, but also by their ability to reliably retain context, retrieve previously acquired information, and reproduce procedural rules within long-term managerial processes. For public organizations, this aspect is particularly critical, as administrative activity relies on stable memory of regulatory requirements, organizational rules, facts of specific cases, and their chronological order.Purpose. The purpose of the article is to determine the limits of applicability of agent systems based on large language models in public administration through a comparative analysis of human memory models and the memory architectures of such systems.Methods. The study employs a comparative analytical approach. Basic cognitive models of human memory are examined, and their characteristics relevant to managerial activity are identified. Subsequently, architectural mechanisms of information storage and retrieval in agent systems based on large language models are analyzed as a functional analogue of human memory.Results. The analysis demonstrates that agent systems reproduce certain external functions of human memory through a combination of short-term contextual representations and external knowledge repositories. At the same time, fundamental differences are identified, including the absence of autobiographical memory, experiential chronology, embedded responsibility mechanisms, and causal verification. These limitations increase the risk of contextual distortion and complicate the validation of generated outputs.Conclusions. It is concluded that, at present, agent systems based on large language models cannot be used for autonomous decision-making in responsible administrative procedures. Nevertheless, they show significant potential as cognitive assistants for public servants, provided that mandatory human oversight is maintained and personal responsibility for decisions is preserved.