Mobile Background Processing for Persistent On-Device Cognitive AI Systems: A Technical Survey

Overview This paper presents a comprehensive technical survey of mobile background processing frameworks and on-device AI inference capabilities, conducted from the perspective of a practitioner who has designed and deployed a production persistent cognitive AI system on iOS. No proprietary system design, unpublished technical implementation, or pending intellectual property is disclosed. All material surveyed is drawn from publicly available platform documentation, developer resources, and open-access research. The Engineering Problem Persistent, always-on cognitive AI systems — on-device AI that maintains continuous user context and identity state across sessions separated by hours or days — require periodic background maintenance operations that cannot be completed during active user sessions alone. Mobile operating systems impose strict constraints on background execution. This survey characterises what is actually possible within those constraints in 2026, rather than what is conventionally assumed. Scope Three interrelated bodies of publicly documented technical material are surveyed in depth: (1) the iOS background task execution framework, covering BGAppRefreshTask, BGProcessingTask, and the BGContinuedProcessingTask introduced in iOS 26, including scheduling heuristics, constraint configuration, preemption handling, and the BackgroundTasks framework evolution from iOS 7 through iOS 26; (2) the Android background processing framework, covering WorkManager, JobScheduler, Doze mode, and App Standby bucket policies, including the historical evolution from AlarmManager and legacy background services to the modern declarative constraint model; and (3) the on-device AI inference landscape under mobile hardware constraints, including battery and thermal characterisation across CPU, GPU, and Apple Neural Engine (ANE) execution pathways, quantisation strategies for memory-feasible on-device model deployment, thermal throttling behaviour under sustained inference load, and the current mobile LLM inference framework landscape including Core ML, MLX, llama.cpp, MLC-LLM, and ANEMLL. Key Findings Mobile platforms in 2026 provide a richer and more capable background execution environment for AI workloads than is commonly assumed. BGProcessingTask with charging-state constraints and the newly introduced BGContinuedProcessingTask together enable meaningful AI maintenance workloads during device idle periods on iOS. Android's WorkManager provides equivalent capability with guaranteed task persistence across device reboots. Both platforms converge on the overnight charging period as the optimal execution window for compute-intensive AI maintenance. At 4-bit quantisation, 1B–3B parameter models are memory-feasible on current iPhone hardware and can execute at 9–62 tokens per second via the Apple Neural Engine at approximately 2 watts — a power budget compatible with charging-period background execution. Design Implications The survey derives a comprehensive set of design principles for persistent cognitive AI systems on mobile: treating background execution windows as first-class architectural primitives; implementing checkpoint-resume patterns for preemption safety; selecting low-power inference pathways for battery conservation; and scheduling maintenance operations around charging-constrained windows rather than deferring them opportunistically. Structure 10 sections · 9 figures · 35 references · ~11,000 words · Compliant with ISO 214, ISO 215, ISO 690, and APA 7th Edition citation standards. Declaration The author declares no competing interests and no funding was received for this work.