The Eleven Open Problems of Multi-Agent Memory Consistency: A Constructive Status Map

A 2026 vision paper argues that multi-agent memory should be read through the lens of computer architecture, and names multi-agent memory consistency as the field's most pressing open challenge: hardware has had half a century of consistency models and coherence protocols with formal verification, and agent memory has no equivalent. This note is a constructive response. Working from an eleven-item decomposition of that agenda, we report, item by item, what a machine-checked semantic-memory substrate does and does not answer - and we sort the answers into four strata that we deliberately never blur: (I) machine-checked under TLC at bounded constants, (II) reduced to invariants already established in that envelope, (III) answered on a different axis than the one on which the problem was posed, and (IV) honestly still open. Two items land in stratum IV and stay there: semantic identity and merge, which we measure as an approximately 40% ceiling that is an information deficit rather than a method deficit, and implicit inter-agent dependencies, where reasoning-only descendants are recovered by no detector we tested. The value claimed here is the map, not the conquest.