Stop Spatializing Time: Machine Learning Agents Should Learn Through Time, Not About Time
Teeratham Vitchutripop, Alyssa Quarles, Wei Zhang, Daniel Rakita
Modern machine learning systems are increasingly deployed in settings that require persistent interaction, adaptation, memory, and decision-making over time. Yet, most learning paradigms remove the temporal pressures faced by physically embedded agents: the world waits for computation, failures are erased by resets, past experience can be replayed exactly, memory is treated as static storage, and learning is often separated from the irreversible trajectory of the learner. We argue that these assumptions limit progress toward agents that can learn through experience over extended lifetimes. This paper proposes six guidelines for temporally grounded learning: autonomous external time, partial embodied observability, non-resettable lifetime interaction, bounded reconstructive memory, multi-rate endogenous computation, and path-dependent self-modification. We use these guidelines to compare major machine learning paradigms, showing that existing approaches capture important fragments of temporal grounding while relaxing, externalizing, or assuming away other temporal pressures. We then present calls to action for the community to build benchmark ecosystems that make temporal grounding measurable, audit the temporal assumptions inside agent architectures, create shared venues around time, and reward temporally grounded work during review. Our central claim is that modeling temporal structure is not sufficient for persistent agency: artificial agents should learn, act, remember, and adapt within the same irreversible time in which their worlds unfold.