Modern AI systems can retrieve, generate, and act, yet they typically lack a governed memory: a persistent substrate that decides not only what to remember, but also what should be trusted, recalled, updated, forgotten, or blocked. Stateless large language models make this gap structural — a fixed knowledge cutoff, no memory across sessions, ungrounded recall, and no intrinsic control over what the system commits to memory or acts upon. We argue that the missing ingredient is not more parameters but governance treated as a first-class operation, so that memory becomes a controlled resource rather than a passive storage layer. We propose a governance-first memory architecture in which every memory transition — write, retrieval, consolidation, forgetting, and self-directed retraining — is mediated by an explicit policy gate and recorded as auditable evidence. The architecture rests on three ideas: a governance model G = (P, A, E) that admits a memory change only when policy is satisfied; a per-item trust model and trust-aware retrieval that rank recall by trustworthiness and utility rather than similarity alone; and a governed cognitive cycle that binds bio-inspired memory layers into corrigible, goal-directed reasoning. The formalism is implementation-independent; we instantiate it in the Matrix OS governed-autonomy operating system as one realization. In a small, reproducible study, governed retrieval enforces source-cited recall and improves governance-decision accuracy over a comparable open baseline at millisecond latency, and a controlled ablation shows that adding the trust signal drives the selection of a plausible-but-poisoned memory item from every query down to none. We present these as evidence that the mechanisms behave as designed, not as a claim of frontier capability. The contribution is an architecture and formalism for making memory inspectable, scoped, and reversible in continual AI systems.
R.I.M. Vsevolodovna (2026) studied this question.