Cross-Device Semantic Memory Persistence: Zero-Cognitive-Overhead Inference via Edge Preloading and Multi-Level Hot Caching

Abstract

This paper proposes the Cross-Device Semantic Memory Persistence (CDSMP) architecture, a six-stage pipeline (Perception–Distillation–Caching– Synchronization–Injection–Feedback) that achieves zero-cognitive-overhead semantic memory continuity across user devices. Core innovations include SAN (Semantic Association Network), TMT (Transitive-Multi-Tier) four-level hot caching, DSS (Delta Semantic Sync) incremental synchronization protocol, dual-path injection (Path A explicit context; Path B hidden state fusion), PAMS (Privacy-Aware Memory Segregation) three-level isolation, and the Adaptive Evolution Engine (AEE) with five mechanisms including memory navigation strategy π_nav and preloading time window θ_window. Experiments on LoCoMo, GSM8K, and SWE-bench-lite show CDSMP improves task completion rate to 94.7% in cross-device scenarios, reducing bandwidth by 89% and context tokens by 87% compared to the Full-Context baseline. AEE ablation confirms π_nav elevates Precision@Preload to 78.3% and reduces WasteRate to 14.2%.