Abstract: Objectives Human mobility diaries, represented as sequences of time–place–purpose events, provide a fundamental basis for understanding human–environment interactions and the dynamics of urban systems. Yet high-quality real-world mobility data are increasingly hard to obtain due to privacy constraints, collection costs, and sampling biases, creating a persistent bottleneck for urban studies and geographic process modeling. Large language model agents offer a new direction by simulating human-like decision making, but what does it take for a diary to be mobility-valid, not merely language-plausible? Specifically, how can an LLM generate activity chains that remain consistent with individual attributes and ground abstract intentions in locations that are both semantically appropriate and accessible? Methods To produce synthetic diaries that are both behaviorally plausible and spatially realistic for downstream urban analysis and policy evaluation, we propose LLM-Driven Human Mobility Generation Agent (MobiA), a travel-diary generation framework built on a dual anchoring mechanism i.e., Sociodemographic Grounding and Spatial Attractiveness. First, to enforce attribute–behavior self-consistency, MobiA mines fine-grained mobility groups from large-scale travel surveys and distills pattern–motivation knowledge into a retrievable knowledge base, which is further validated and updated through contextual evaluation. Second, to ensure geographically valid grounding, the city is represented as road-enclosed parcels, and an attractiveness model maps travel intentions to feasible parcels based on semantic supply and accessibility. Conditioned on an agent profile, MobiA performs hierarchical schedule planning and event refinement under retrieved constraints, and then grounds each event to parcel locations via the learned attractiveness scores. Results MobiA is trained and validated using a large-scale Shenzhen WeChat travel survey dataset, which contains 22,228 users, 199,380 trips, and 32,727 locations after preprocessing. (1) We conduct two controlled cohort experiments to test whether MobiA produces attribute-grounded personalization. For women aged 30–35 on weekday mornings, MobiA shifts destinations from a commute-only business-district cluster to increased school-area activity, consistent with family–work routines. For male car owners aged 30–35 with middle-to-high income on weekends, MobiA shifts trips from near-home leisure to longer-distance destinations such as parks, museums, seaside areas, and major shopping centers, demonstrating fine-grained demographic heterogeneity in generated diaries. (2) City-scale realism. We simulate ~50,000 residents and compare 24-hour parcel-level dynamic population against Baidu Maps. As shown in Fig.3, the cosine similarity exceeds 0.8 in >90% of parcels (weekdays and weekends), and >70% of parcels achieve density-fitting >0.5, indicating strong alignment of spatiotemporal population patterns at the parcel scale. Conclusions MobiA generates high-quality, controllable travel diaries with strong behavioral realism and geographic validity. Its dual grounding mechanism links sociodemographic attributes to daily routines and links urban semantics and accessibility to destination choices. We emphasize that individual mobility synthesis must respect both behavioral regularities and spatial feasibility to avoid plausible-butfloating diaries.