Large Language Models don’t actually remember anything yet we keep building AI applications as if they do.

In real-world systems, memory is not a model feature but an architectural responsibility. As soon as we add user context, conversation history, or long-term state, our AI systems start behaving in unexpected ways: recalling outdated information, mixing contexts, or confidently acting on stale data.

In this talk, I’ll walk through how memory really works in modern AI applications, and why naive approaches-like endlessly growing prompts or unstructured context injection, eventually fail.

Using practical Python-based examples, we’ll explore different types of memory (short-term, long-term, and time-aware memory), common failure modes, and design patterns that keep AI behavior predictable and safe over time. I’ll also show how forgetting, summarization, and time-based decisions are just as important as remembering.

This talk is aimed at engineers building real AI products who want their systems to stay coherent, reliable, and maintainable.