El lunes 8/06/26 tendremos una nueva sesión del seminario, con la presentación de Mariano Consens.
Título: Evaluating Memory in LLM Agents (charla en español con slides en inglés)
Lugar: sala de seminarios del INCO
Fecha y hora: Lunes 8/6/2026, 16hs.
Resúmen: As large language model (LLM) agents are deployed in increasingly long-horizon settings—multi-turn dialogues, software development, scientific discovery—a single context window is far too small to capture what has happened, what was learned, and what should carry forward. Memory, the ability to persist, organize, and selectively recall information across interactions, is what turns a stateless text generator into a genuinely adaptive agent.
This talk explores the rapidly evolving landscape of agent memory and its evaluation. We examine the life cycle of memory—extraction, storage, retrieval, and evolution—with a focus on graph-based representations that model relational dependencies and support efficient recall. We then trace the shift from static recall benchmarks to multi-session agentic tests that interleave memory with decision-making, highlighting MemoryArena, a recent benchmark on which agents near-saturated on long-context tests like LoCoMo perform poorly once memory must guide sequential action.
Acerca del expositor: Mariano Consens research interests are in the areas of Data Management and the Web, with a focus on graph data, machine learning and analytics, semantic data, searching, and autonomic systems. He has over 100 publications, including journal publications selected from best conference papers and several patents. Mariano received his PhD and MSc degrees in Computer Science from the University of Toronto, and a Computer Systems Engineer degree from the Universidad de la Republica, Uruguay. Consens is a University of Toronto faculty member. He has been a Visiting Scientist at the IBM Center for Advanced Studies in Toronto and at Yahoo! Research in Barcelona. In addition, he has been active in the software industry as a founder and CTO of a couple of software startups.
Más información en la página del Seminario del Instituto de Computación