Governed Mnemonic-Semantic Support in Longitudinal Human-AI Collaboration

Longitudinal human-AI collaboration creates a problem that is not solved by storage alone. A project may preserve extensive external memory and still require its human participant to reconstruct meanings, distinguish near-neighbour concepts, recover operational constraints, and reconnect an AI successor with the working history that makes those distinctions usable. This paper develops a case- derived account of one response to that problem: a governed mnemonic-semantic layer of Cognitive Supports. The proposal emerges from Dufour 24, a longitudinal human-AI sociotechnical project originally organized around the technical reconstruction of a classic sailing yacht and later also used to study continuity, external memory, authority, documentary repair, and successor readiness. Across that work, a Functional Nomenclature developed alongside short formulations, acronyms, graphic distinctions, metaphors, and contextual humor. These representations were not treated as sources of truth or authority. Their function was cognitive: to compress, orient, retrieve, and discriminate knowledge while keeping the governing meaning, evidence, state, and authority in competent sources. The paper advances three proposed contributions. First, it articulates the integrated architecture of a governed mnemonic-semantic layer in which cognitive usefulness is explicitly decoupled from epistemic or constitutive authority. Second, it presents the Operational Glossary Nomenclature as the most developed operational artifact of that idea: a read-only, version-bound interface for AI currents that assembles meaning, relations, limits, provenance, and application remissions without becoming an authority of its own. Third, it documents a naturalistic longitudinal case in which some co-constructed humorous and metaphorical frames emerged during non-humorous technical and methodological taskwork and later became reusable mnemonic-semantic supports. Two hypotheses remain deliberately open. A previously shared humorous frame may provide a bounded probe of common-ground continuity after interruption, clean restart, or AI succession; and high-salience Cognitive Supports may work best when they remain contextual and proportionate rather than quota-driven. Neither hypothesis is treated as established causal evidence. The paper positions these claims against prior work on external cognition, human-AI co-learning, lexical alignment, jargon support, notation, decision support, humor in task- oriented HCI and human-autonomy teams, and common ground. The claimed contribution is therefore not the invention of any single component, but a specific integration, governance boundary, operationalization, and documented field history.