Abstract The Cognitive Memoisation (CM-2) Normative Architecture [note 1] defines a collection of normative invariants that, together with CM-2 protocol invariants, realise a portable architectural mechanism for enforcing Attention preservation within LLM-mediated inference systems. The architecture addresses the Governance Attention Axis as a primary UN normatively enforceable concern. The central thesis of the architecture is that projected normative Reference Object Collections (ROC) of Reference Objects (RO), governed by normative consistency guarding, provide deterministic detection of Attention deficit, where (in CM-2) Attention is defined as participation in Inference. Incoherence - such as the ejection of Epistemic Objects (EO) or other required artefacts from Inference - is treated as a precursor to phenomena commonly described as drift or forgetfulness. Participation in Inference is a precondition for epistemic effect. When required epistemic artefacts fail to participate in Inference, their influence becomes null. In contemporary LLM systems, such exclusion may occur silently due to truncation, salience pressure, summarisation, probabilistic reconstruction, or context eviction - leading to drift, expulsion, or authority inversion. CM-2 transforms this silent exclusion into an explicit, normatively detectable violation, which is then deterministically remediated. While improvements across other Governance Axes may arise as a consequence of sustained Attention preservation, this architecture normatively enforces only the conditions required for deterministic detection and remediation of Attention deficit within single-session scope. Such artefacts may include RO, EA, EO, governed knowledge, constraints, scope declarations, or any other normatively relevant material required for coherent Inference.