Organizations increasingly want AI agents that retain context over time. However, building "agentic memory" requires moving beyond simple language models to complex systems that must autonomously extract, synthesize, and retrieve information. Because these systems are fundamentally non-deterministic, they introduce unique operational hurdles.

This talk explores the current landscape of agentic memory, covering its theoretical promise, common implementation approaches, and practical challenges—such as compounding errors, multi-user conflicts, and the loss of data provenance. Drawing on a decade of data engineering experience, the session offers a pragmatic guide on navigating these systems, highlighting when to invest in emerging frameworks and when to rely on simpler, observable alternatives.