Anna Lisa Gentile

(Senior Research Scientist, IBM)

Bio: Anna Lisa Gentile is a senior research scientist at IBM Research, Silicon Valley Lab, USA. Her main Research Areas are Information Extraction (IE), Natural Language Processing (NLP) and Semantic Web.
Prior to IBM she has been a post-doctoral research associate at the University of Sheffield and a post-doctoral research scientist at the University of Mannheim, anna Lisa obtained doctoral degree with a thesis on Named Entity Disambiguation at the University of Bari, Italy in 2010.

Title of Talk: From Conversation to Memory: Consolidating What Agents Learn About Us

Abstract: AI agents are quickly becoming the primary interface to our digital lives, and in the process they accumulate something genuinely new: a private, volatile record of who we are, assembled from conversation rather than from forms and databases. This conversational memory is much of what makes an agent feel personal, and it is also what makes it risky. Memory that cannot be inspected, corrected, or attributed turns helpfulness into guesswork.
In this talk I will argue that the hard problems in agentic memory are not primarily retrieval problems but representation and governance problems: how knowledge is structured as it is consolidated, how it is revised when a user changes their mind, how provenance is preserved back to the utterance that produced it, and how it is used at query time without being paraphrased into something the user never actually said.
Drawing on our work in this space, I will make the case for Knowledge Graphs as the substrate for agentic conversational memory, paired with a mechanism that combines imperative and generative computing to produce trustworthy, traceable responses. Rather than trading determinism for fluency, the two are separated: the graph determines what is known, the model determines how it is expressed. I will close with lessons from our enterprise pilots – where structure paid off, where it got in the way, and the open questions for this community: conflict across time and agents, staleness, and scoping. 

Bryan Perozzi

(Research Scientist, Google Research)

Bio: Bryan Perozzi is a Research Scientist in Google Research, where he routinely analyzes some of the world’s largest (and perhaps most interesting) graphs. Bryan’s research focuses on developing techniques for learning expressive representations of relational data with neural networks. These scalable algorithms are useful for prediction tasks (classification/regression), pattern discovery, and anomaly detection in large networked data sets. Bryan is an author of 40+ peer-reviewed papers at leading conferences in machine learning and data mining (such as NeurIPS, ICML, ICLR, KDD, and WWW). His doctoral work on learning network representations (DeepWalk) was awarded the prestigious SIGKDD Dissertation Award. Bryan received his Ph.D. in Computer Science from Stony Brook University in 2016, and his M.S. from the Johns Hopkins University in 2011.

Title of Talk: A Few Foundational Takes on Connected Data

Abstract: Graphs are powerful tools for representing complex real-world relationships, essential for tasks like analyzing social networks or identifying financial trends. While large language models (LLMs) have revolutionized natural text reasoning, their application to graphs remains an understudied frontier. In this talk, I’ll describe some highlights from our journey making LLMs more relational, and then focus on our new work making graph models more foundational. I’ll introduce the Universal Classifier (UC), which supports arbitrary feature and class cardinalities spanning all learning objectives through the in-context learning paradigm. The UC reformulates all node-, edge-, and graph-level prediction tasks as maximizing similarity in the latent space: by lifting heterogeneous features and labels into 3D latent tensors, the model learns transferable features independent of specific input schemas. This architecture allows a single pre-trained model to generalize to node classification, node regression, and link prediction across unseen graphs with varying feature semantics. Experiments show strong zero-shot transfer performance across node- and link-level tasks.