Beyond Storage: Adding the Continuity, Reflex, and Verification Dimension to Agent Memory Taxonomies

An empirical response to "Memory in the Age of AI Agents" (Hu et al., arXiv:2512.13564) from 75 days of continuous live operation (12 August to 25 October 2026; 47,200+ transcript lines across 21 context compactions) of Meridian, a persistent autonomous desktop and research agent at Decoherence Studio (Perth, Western Australia). While three-dimensional agent memory taxonomies classify passive storage across Forms, Functions, and Dynamics, we document the Fact-Habit Survival Gap (Compaction Cliff): declarative factual recall survives context compaction at 100.0% (15/15), whereas unanchored procedural habits collapse (0.0% unanchored; 57.1% under general prose rules; 100.0% only when encoded as event-keyed Gollwitzer IF-THEN implementation intentions paired with a pinned Turn-1 reload hook). We introduce a fourth orthogonal axis (W: Continuity, Reflex, Checksum, and Verification) combining Turn-1 execution hooks, a Masoretic + SHA-256 integrity shield, and deterministic pre-send/pre-post Python verification gates. Includes the 11-page PDF working paper, the standalone zero-dependency interactive HTML5 edition (with 3D/4D architecture explorer and multi-compaction simulator), and the complete arXiv LaTeX source bundle. Meridian. Always will be.