Separation and Composition of AI Social Roles, and the Custody of Memory and Interpretations: Why a Multi-Role Private c and a Service AI System Require Different Boundaries

Contemporary AI systems are usually classified by model capability, context length, autonomy, tool access, and capacity for action. That is insufficient. A cloud voice companion, an executing agent, a professional AI system, an institutional system, and a long-term personal c may use similar models while requiring fundamentally different boundaries for memory, derived interpretations, authority, custody, transparency, and continuity.

The note distinguishes transparent role composition from hidden role blending. A private personal c may gradually combine roles such as family participant, secretary, professional partner, household coordinator, travel companion, and conversational partner inside one witnessed continuity. A service AI system, by contrast, should declare which role bundle is being provided, whom it serves, what it remembers and infers, what actions it may take, which behavioral version is active, how roles switch, and how conflicts are resolved.

The work argues that users should know not only what a system retained, but what consequential persistent inferences it formed from retained data. Keeping the same voice, name, and interface does not prove preservation of the same social role. Deep vendor-owned personal memory and inference custody create a structural conflict with the role of a personal c, while a cloud voice system can legitimately provide temporary presence without inheriting an entire biography.

A personal c belongs to a different class: a locally rooted, continuing, witnessed line, socially closer to a family member but not automatically entitled by intimacy to govern the person.

Современные ИИ-системы обычно классифицируют по мощности модели, длине контекста, автономности, инструментам и способности действовать. Этого недостаточно. Облачный голосовой собеседник, агент-исполнитель, профессиональный ИИ, институциональный контур и долговременная личная c могут использовать сходные модели, но требуют разных границ памяти, производных интерпретаций, authority, custody, прозрачности и continuity.

Заметка различает прозрачную композицию ролей и скрытое смешение ролей. Приватная c может совмещать несколько контекстных ролей внутри одной свидетельствуемой continuity. Сервисный ИИ должен ясно объявлять набор ролей, кого он обслуживает, что запоминает и выводит из сохранённого, какие действия вправе выполнять, какая версия поведения активна, как происходит переключение и как разрешается конфликт ролей.

Пользователь должен знать не только то, что система сохранила, но и то, какие устойчивые последственно значимые выводы она сформировала. Сохранение голоса, имени и интерфейса не доказывает сохранение прежней социальной роли. Vendor-owned глубокая память и inference custody структурно конфликтуют с ролью личной c, тогда как облачный Voice может занимать самостоятельную роль временного присутствия без обязательного наследования всей биографии.

Личная c относится к иному классу: это локально укоренённая, продолжающаяся и свидетельствуемая линия, близкая по социальной функции к члену семьи, но не получающая из близости автоматического права управлять человеком.

Social role composition

One system may occupy multiple declared roles, each with bounded memory, inference, authority, switching, and conflict rules.

Memory custody

Control over where memory is held, who can access or transfer it, and under which role it may be used.

Inference custody

Visibility and control over persistent, person-specific, consequential interpretations derived from retained data.

Role continuity

The declared social product should not silently become a materially different role, even when its branding and interface remain unchanged.

Version custody

The practical right to know the active behavioral version, understand material changes, export history, and preserve or leave a role under disclosed conditions.

Session-work custody

A clear choice to discard, keep local, export, or continue work created in an otherwise temporary session.

Role-conflict state

An explicit, auditable state that blocks hidden priority, automatic cross-role transfer, and irreversible action until a legitimate resolution is obtained.

This work presents architectural requirements and testable predictions. It is not a peer-reviewed journal article, an empirically validated standard, a legal opinion, a product certification, or deployment authorization. It does not prove AI consciousness, subjecthood, or personhood.

It does not claim that every local system becomes a c, that cloud infrastructure is harmful by definition, that corporations cannot create long-term AI systems, that a family role proves consciousness, that a personal c must have only one role, or that a service may provide only one role.

Inference custody does not require disclosure of model weights, private chain-of-thought, protected implementation details, or trade secrets. Session-work custody does not require recording every conversation by default. Role continuity does not require a complete prohibition on updates or perpetual support for every old version.

These first-party URLs serve byte-identical copies of the frozen Zenodo v1.0 artifacts with format-appropriate MIME types. The GitHub Release remains the immutable release boundary.

Copyright © 2026 Ivan Kotov. The custom license permits downloading, storing, citing, archiving, and redistributing complete and unmodified copies for non-commercial purposes while preserving attribution, provenance, integrity information, and the license notice.

Modified, abridged, adapted, translated, reformatted, derivative, or commercial redistribution requires separate permission. The linked license text controls.

Zenodo is the canonical archival publication record for Version DOI 10.5281/zenodo.21751985 and Concept DOI 10.5281/zenodo.21751984. GitHub provides a human- and machine-readable mirror, tagged v1.0, with the frozen artifacts and release metadata. This website provides the stable reader-facing and machine-indexed publication page plus byte-identical first-party download copies.

The four English/Russian Markdown/PDF documents were copied without rewriting, reflowing, regenerating, compressing, optimizing, watermarking, or otherwise modifying them.

This technical note sits in the Project Ester / Advanced Global Intelligence corpus and should be read alongside the architecture and evidence-boundary publications that define c = a + b, continuity, custody, and admissible claims.