Memory, Meaning, and Machines: Building the Knowledge Scaffolds of Agentic AI
Abstract
Agentic AI systems promise autonomy, adaptivity, and natural interfaces, but they struggle with grounding, memory, and maintaining an evolving understanding of the world. In this talk, I argue that structured knowledge (particularly in the form of knowledge graphs) provides the scaffolding agents need to operate reliably at scale. Leveraging more than two decades of progress in querying, reasoning, and knowledge engineering, we already have powerful tools that complement modern LLMs rather than compete with them. I will illustrate this through GraphRAG and examples from large-scale multimodal and socio-political KGs. Through these case studies, I will highlight both the possibilities and the engineering challenges ahead: scalable graph and vector systems, evaluation of grounded agents, multimodal schema design, and agent architectures that can read, write, and evolve structured knowledge.
