(Abstract + short pitch) Generative AI models (LLMs and diffusion/video models) achieve state-of-the-art results in isolated tasks but suffer from Contextual Fragmentation in temporal workflows, leading to continuity drift and expensive human rework. This paper introduces the Mnemosyne Protocol, a vector-based orchestration layer designed to maintain semantic and visual consistency across heterogeneous generative systems while enforcing Local-First Sovereignty for studio IP. Mnemosyne operationalizes continuity as a discrete-time, fail-closed verification gate (a conjunctive product-of-constraints) evaluated across frame sequences, with an explicit rollback + localized re-sampling mechanism before final rendering. Preliminary simulations demonstrate substantial reductions in continuity hallucination rates under defined constraints. What’s new in v1.5:v1.5 introduces a discrete-time, fail-closed continuity gate (a conjunctive product-of-constraints), adds formal definitions/propositions, and specifies an explicit rollback + localized re-sampling algorithm for continuous verification during generation. Updates are additive and backward-compatible at the protocol level. Code is licensed under MIT (see LICENSE in the archive); the paper text is CC BY 4.0.
Mert Kerem Salman (Sun,) studied this question.