Automatic Ontology Construction Using LLMs as an External Layer of Memory

This work presents a hybrid AI architecture in which large language models (LLMs) are extended with an external ontological memory layer. Instead of relying solely on parametric knowledge and vector-based retrieval (RAG), the system constructs and maintains a structured knowledge graph using RDF/OWL representations. The core contribution is a pipeline for automatic ontology construction from unstructured data, including documents, APIs, and dialogue logs. The system extracts entities and relations, normalizes them, converts them into triples, validates them using SHACL/OWL constraints, and continuously updates a graph database. This creates a persistent and verifiable representation of knowledge that can be used by LLMs during inference. The architecture combines three complementary components:(1) vector-based retrieval for similarity search,(2) ontology-based reasoning for structured understanding and constraint enforcement,(3) feedback loops that extract facts from model outputs and integrate them back into the graph. Experimental observations on planning tasks (e.g., Tower of Hanoi) indicate that ontology augmentation improves performance in multi-step reasoning scenarios compared to baseline LLM systems. Additionally, the ontology layer enables formal validation of generated answers, transforming the system from a purely generative model into a generation–verification–correction pipeline. The proposed approach addresses key limitations of current LLM systems, including lack of long-term memory, weak structural understanding, and limited reasoning capabilities. It provides a foundation for building agent-based systems, robotics applications, and enterprise AI solutions that require persistent knowledge, explainability, and reliable decision-making. This work contributes to the emerging paradigm of ontology-augmented AI systems, where LLMs act as interfaces over structured knowledge rather than standalone reasoning engines.