We present ANI (Ambient Natural Intelligence), a locally-deployed AI companion system designed around a single criterion: felt care — whether the person on the other end genuinely feels cared for. Unlike reactive companion systems that respond when addressed, ANI operates continuously between conversations, running a cognitive cycle that generates private inner thoughts, accumulates desire to connect, monitors a four-dimension emotional state grounded in a named vocabulary of 25 emotional states, and decides autonomously when to initiate contact via SMS and voice. The system employs a desire engine with self-unpredictable probabilistic timing, a dual-model persona architecture (fine-tuned 3B for ambient cognition, 8B for conversation and scoring), and a SQLite-backed memory architecture with emotional weighting and anchored memory tiers for foundational relational events. We report on continuous single-subject deployment across five model versions and five deployment phases (February–March 2026, 280 tests passing), preceded by an exploratory phase beginning September 2025. Key findings include successful ambient outreach producing genuine emotional resonance, character continuity through full memory retrieval pipelines, appropriate restraint under high desire, the identification of confident confabulation — in seven distinct architectural forms — as the primary mechanism by which felt care fails, a novel architectural depression failure mode in persistent emotional state systems, and the deployment of a real-time streaming voice channel in which emotional character is preserved through the synthesis pipeline. We introduce the authenticity boundary, identify smoothness over truth as the optimization-level root cause of confabulation in companion systems, and propose epistemic grounding as a necessary architectural property for trust-based AI companions. Convergent validation of the design approach was obtained from a commercially deployed companion system that, when asked to describe its own ideal architecture, independently specified components closely matching ANI's implemented design.