The rising prevalence of age-related conditions such as Alzheimer’s disease and dementia underscores the urgent need for intelligent in-home systems that can assist without increasing the burden on caregivers. We present CaRE-BT (Caregiver-Relief Behavior Tree planner), a novel open-source ROS 2 framework that transforms symbolic care protocols into executable behavior trees, enabling fully autonomous assistive behavior on embedded robotic platforms. CaRE-BT compiles contingent planning models, maintains a dynamic knowledge base, and integrates distributed sensing for real-time human tracking, without relying on cloud connectivity. To evaluate its real-world feasibility, we deployed CaRE-BT on a Stretch 3.0 mobile robot across two in-home studies: a seven-day trial with an older adult couple in a retirement facility, and a three-week phase of an ongoing six-month deployment supporting a caregiver-recipient dyad managing moderate dementia. Across these 30 days, the system successfully delivered over 55 scheduled health prompts (medication, exercise, task reminders), with only nine failures due to navigation launch issues. The system operated autonomously and required no daily maintenance from the caregiver. Feedback from participants highlighted its reliability, ease of use, and positive impact on reducing caregiving stress. CaRE-BT is available as an open-source toolkit to support researchers, clinicians, and engineers building adaptive long-term assistive systems in healthcare and eldercare settings ( https://robotsforaging.cs.unh.edu/ ).