The research of artificial intelligence is undergoing a paradigm shift from prioritizing model innovations over benchmark scores towards emphasizing problem definition and rigorous real-world evaluation. As the field enters the “second half,” the central challenge becomes real utility in long-horizon, dynamic, and user-dependent environments, where agents face context explosion and must continuously accumulate, manage, and selectively reuse large volumes of information across extended interactions. Memory, with hundreds of papers released this year, therefore emerges as the critical solution to fill the utility gap. In this survey, we provide a unified view of foundation agent memory along three dimensions: memory substrate (internal and external), cognitive mechanism (episodic, semantic, sensory, working, and procedural), and memory subject (agent- and user-centric). We then analyze how memory is instantiated and operated under different agent topologies and highlight learning policies over memory operations. Finally, we review evaluation benchmarks and metrics for assessing memory utility, and outline various open challenges and future directions.

Researchers have significantly enhanced an artificial intelligence tool used to rapidly detect bacterial contamination in food by eliminating misclassifications of food debris that looks like bacteria. Current methods to detect contamination of foods such as leafy greens, meat and cheese, which typically involve cultivating bacteria, often require specialized expertise and are time-consuming—taking several days to a week.

Luyao Ma, an assistant professor at Oregon State University, and her collaborators from the University of California, Davis, Korea University and Florida State University, have developed a deep learning-based model for rapid detection and classification of live bacteria using digital images of bacteria microcolonies. The method enables reliable detection within three hours. The findings are published in the journal npj Science of Food.

Their latest breakthrough involves training the model to distinguish bacteria from microscopic food debris to improve its accuracy. A model trained only on bacteria misclassified debris as bacteria more than 24% of the time. The enhanced model, trained on both bacteria and debris, eliminated misclassifications.

A team of EPFL researchers has taken a major step towards resolving the problem of drift in generative video, which is what causes sequences to become incoherent after a handful of seconds. Their breakthrough paves the way to AI videos with no time constraints.

A fundamental desideratum of AI is the ability to model environment dynamics and transitions in response to both their own actions and external control signals. This capability, commonly referred to as world modeling (WM), is essential for prediction, planning, and generalization. Learning world models using deep learning has been an active area of research for nearly a decade. In recent years, the field has witnessed significant breakthroughs driven by advances in deep neural architectures and scalable learning paradigms. Multiple subfields, including self-supervised learning (SSL), generative modeling, reinforcement learning (RL), robotics, and large language models (LLMs), have tackled aspects of world modeling, often with different tools and methodologies. While these communities address overlapping challenges, they frequently operate in isolation. As a result, insights and progress in one area may go unnoticed in another, limiting opportunities for synthesis and collaboration. This workshop aims to bridge this gap between subfields of world modeling by fostering open dialogue, critical discussion, and cross-disciplinary exchange. By bringing together researchers from diverse backgrounds, from early-career researchers to established experts, we hope to establish a shared vocabulary, identify common challenges, and surface synergies that can move the field of world modeling forward.