Noa Barbiro

As LLMs become increasingly capable, a core limitation is no longer raw intelligence, but how systems decide what information to use, retain, and forget over time. In real‑world settings, especially knowledge‑intensive work, AI must operate across extended interactions, evolving goals, and heterogeneous data sources. Simply increasing context windows or replaying full histories is neither scalable nor reliable.

This shift has brought memory and context engineering to the forefront of modern AI system design and key driver for personalization. Rather than treating context as a static prompt or memory as a passive log, emerging agentic systems actively manage what information persists, what is retrieved on demand, and how past interactions shape future reasoning. These design choices have profound implications for response quality, reliability, bias, interpretability, and user trust.
This roundtable will explore how agentic memory and context selection are reshaping the way we build, evaluate, and reason about intelligent systems, with a focus on principles that generalize across domains, organizations, and platforms.

Key Questions for Roundtable Discussion-

Participants will be invited to discuss and debate questions such as:
* What does it mean for an AI system to “remember” responsibly? How should systems balance persistence, relevance, freshness, and forgetting?

*Context as a scarce resource: How do we decide what information is worth injecting into a model at a given moment, and what should stay latent or external?

* Agentic vs. static memory: When should systems actively update or restructure memory versus relying on fixed retrieval pipelines?

* Data science & evaluation implications: How do memory and context decisions affect evaluation, bias amplification, error compounding, and longitudinal performance?

* Measurement & accountability: What metrics or evaluation frameworks help us understand whether memory is helping or degrading response quality over time?