The Great Amnesia: AI Agent Memory as the Binding Constraint on Autonomous Systems in Enterprise and Defense

Abstract

Frontier model capability has improved rapidly for four years while the rate at which organisations convert artificial intelligence pilots into production has not. This paper argues that the two facts are connected, and that the binding constraint is no longer reasoning quality but continuity: what an autonomous system carries from one operating cycle to the next. Four independent lines of published evidence are assembled. A frontier laboratory's own usage data shows computer and mathematical tasks at roughly 35 percent of consumer conversations and close to 44 percent of first-party API traffic, with software error correction the single most frequent task. Enterprise research attributes an approximately 95 percent pilot failure rate not to model quality but to a learning gap. Time-horizon measurement shows long 50 percent horizons alongside much shorter 80 percent horizons. And long-horizon agent performance degrades sharply when a task is embedded in a longer interaction history even while the required information remains inside the context window, which locates the failure in architecture rather than in information availability. The paper then examines the same pattern in defense programmes, argues that stored facts, larger context windows and similarity-based retrieval are each insufficient for continuity, specifies eight properties a continuity layer must exhibit, proposes a falsifiable two-cycle evaluation protocol, and states four checkable predictions. This is a position and architecture paper. It presents no original empirical work and no validation of the author's own systems. The author declares a competing interest as co-founder and Chief Executive Officer of Rebootix AI, Inc. Companion preprint: 10.5281/zenodo.22557263.