Long-term dementia care is often shaped by fragmented documentation practices that fail to capture the relational and evolving nature of caregiving. We present CareInsights, an AI-enabled infrastructure that supports caregivers by passively capturing interaction data, enriching it with contextual and clinical signals, and generating interpretable summaries to support day-to-day care. Our system was developed through a two year collaboration with long-term care facilities and integrated into existing health record platforms to minimize disruption and streamline uptake. We evaluated CareInsights through a combination of system performance analysis, structured content assessments, caregiver surveys, and interviews with other stakeholders. Caregivers reported improved access to meaningful information, reduced documentation burden, and greater confidence in care planning. Adoption was supported by the system's transparency, narrative format, and alignment with established routines. We reflect on design considerations for embedding AI into real-world care environments and highlight how infrastructure-level decisions shape the feasibility and utility of such tools in practice. Please see the attached PDF for more information.