Modeling a user's evolving goals, values, and affect over time is central to perspective-aware AI, yet progress is bottlenecked by the lack of longitudinal data with ground-truth labels for latent identity state. We introduce PAiNT (Perspective-Aware AI Identity and Narrative Toolkit), a generative framework that simulates long-horizon persona trajectories and emits corresponding multimodal artifacts with ontology-aligned labels of the latent identity state that produced them. PAiNT decouples identity dynamics from artifact generation via a typed Persona Matrix and Situation Graph, coordinated through a multi-agent loop with validation-gated transitions and bounded-window history conditioning. Across four personality archetypes, four backbone LLMs, and three architectural ablations, evaluated with a nine-metric suite calibrated on published longitudinal data, we find that (i) persona initialization produces a durable identity signal that persists above stochastic event noise; (ii) multi-agent orchestration and history conditioning govern distinct quality dimensions, with removal of either causing different failure modes; and (iii) a coherence frontier constrains the trade-off between temporal resolution and horizon, with substantial penalties at daily granularity. We release PAiNT and PAi-Bench, a human-validated benchmark of 1,200 labeled multimodal artifacts, at: https://anonymous.4open.science/r/paint-0411/.