Agents have advanced from primarily single-session prototypes to being able to solve long-horizon tasks, but their memory is still commonly implemented as accumulated chat history or basic retrieval. There is no settled architecture for deciding what an agent should store, retrieve, revise, and forget. This track is for researchers and senior engineers building agents for long-running tasks, recurring workflows, or use across multiple sessions.

Talks cover working, episodic, semantic, and procedural memory; memory-writing, retrieval, consolidation, and forgetting policies; context management, external stores, and learned memory; evaluation of recall, relevance, consistency, and task performance; and stale memories, conflicting information, privacy, and permissions. Expect concrete architectures, evaluation results, and discussions about failure modes. Attendees will leave with a stronger understanding of how to choose, implement, and evaluate memory components in an agent system.