Research Paper Discussion #2

A discussion on memorization in models and agentic memory. We explored how memorization experiments show that models can learn even random noise and still achieve 100% training accuracy—and what that tells us about learning versus generalization.

We discussed how models tend to learn the easier, meaningful patterns first before gradually memorizing harder or noisy samples, drawing a parallel with a school curriculum: start with simple concepts, then introduce complexity. We also explored how the time a model takes to learn different samples can reveal something about their complexity.

The second part focused on agentic memory—how AI agents should store, retrieve, and manage information over time to make their memory useful rather than simply accumulating everything.