Mini Me: Personal AI That Was There Continuum, a Context Substrate for Measurable Presence and Safe Action

<jats:p>Today’s personal AI can read an inbox, summarize a meeting it joined, or act inside one application. It still loses the thread when a task crosses a conversation, a screen, a room, the body, and several devices. The person must then reconstruct that thread before asking for help. We argue that the missing capability is _presence_: the system was not there when the context was made, so the person must stand in for it later. _Mini Me_ names the experience of a personal AI that was present across a person’s conversations, screens, body, devices, and spaces. _Continuum_ is the context substrate we propose beneath that experience: a set of person-controlled device trust domains that capture across devices, distill typed memory, compile a purpose-scoped disclosure bundle for any remote processing, and gate proactive action by risk. We define presence so that it can be measured and falsified, along four axes: _continuity_, policy-permitted context coverage by modality, device, and time; _comprehension_, briefing elimination and time to useful context; _anticipation_, next-action precision and useful lead time; and _authority_, autonomy yield bounded by regret, correction, and undo. Presence is deliberately bounded, freshness-aware, and policy-permitted; it is not a euphemism for total capture. The paper’s narrow novelty claim is the proposed composition of measurable presence, typed and provenance-aware memory, purpose-scoped disclosure, cumulative modeled exposure, and content-independent authority. It contributes a reference architecture with a presence fabric, plane contracts, and per-device trust domains inside one personal trust perimeter; a proposed autonomy model with asymmetric calibration, signed grants, provider-verified action preconditions, engineered reversibility, an attention-aware delivery policy, and a four-gate action decision procedure; and an evaluation agenda. All of it is a falsifiable proposal, not an implemented system. Local storage does not defeat endpoint compromise; disclosure metadata cannot force a remote processor to forget; always-available sensing remains socially and legally dangerous. We state these limits directly and propose the experiments that would confirm or refute the design.</jats:p>