Background: Alzheimer disease and age-related cognitive decline reduce memory engagement and limit caregiver insight, creating a need for accessible tools that support everyday cognitive activity in older adults. Although AI holds promise for personalized cognitive support, few AI-based apps have been developed and evaluated for memory engagement in this population, and fewer incorporate on-device emotional analysis with privacy-preserving design. Objective: This study developed and evaluated RecallLive, an AI-assisted mobile app supporting structured memory interactions for older adults, and examined its #usability, engagement, perceived usefulness, and behavioral intention, guided by the Technology Acceptance Model. Methods: RecallLive integrates metadata-based photo clustering to generate memory videos, an on-device convolutional neural network that classifies frame-level emotional responses, and a large language model (LLM)–based module that converts these outputs into caregiver summaries. A sequential 2-phase mixed methods design was used. In phase 1, 202 US adults aged 65 years or older viewed a structured demonstration and completed a survey measuring ease of use, engagement, design clarity, perceived usefulness, and intention to use. In phase 2, 10 participants completed a hands-on session followed by semistructured interviews analyzed thematically. Quantitative analyses were conducted using Python (v.3.11; Python Software Foundation) and included internal consistency estimates; 2-tailed, 1-sample tests against the scale midpoint; and simple linear regression. Results: All subscales showed strong reliability (Cronbach =0.86-0.92). Ease of use (mean 3.77, SD 0.78) and engagement (mean 4.09, SD 0.67) significantly exceeded the scale midpoint (=14.00 and =23.19; both