From Memory to Agency.

This paper explores AI agent memory as a governance problem rather than a storage problem. Its starting point is that preserving a conversation or retrieving past information is not enough to create usable experience. Once past information can influence a present decision, memory acquires operational power. The question then becomes not only what an agent remembers, but what past experience is allowed to influence. The paper distinguishes traces, episodes, learning and recall, and examines mechanisms for preserving experience without silently turning past interpretations into current authority. The proposed approach combines sourced episodes, provenance, revisable learning, quarantine, revocation without deletion, bounded recall and contextual memory packs. A central principle is the separation between preservation and authorization. A belief can remain part of the historical record while losing the right to guide a future decision. Similarly, retrieved information should not automatically become directive context or justify an automated action. The paper also examines the decision boundary between memory and action, with the goal of making observable what was retrieved, admitted, rejected, presented and eventually authorized. Cortex is the experimental implementation used to explore this architecture. The paper explicitly distinguishes mechanisms already materialized in the system from those that still require validation in real-world use. It concludes with a validation program covering memory revocation, unjustified generalization, the measurable influence of recalled artifacts on decisions, feedback from action outcomes into future learning, and the computational and contextual cost of governed memory.