Long-Term Memory Construction and Cognitive Evolution of an Autonomous Agent in a Restricted Mobile Environment: A Case Study of Localized DSH Deployment on Android

This study reports an eight-day, 360-turn practice of long-term memory construction and cognitive evolution by a single autonomous agent on a permission-restricted OPPO education tablet (8 GB memory, no accelerator, remote foundation-model inference). Unlike on-device deployment studies focused on inference efficiency, it asks a more fundamental question: when memory cannot reside in the context window and the runtime may be reclaimed at any moment, on what does an agent's continuity rest? Methodologically, the study is a single-case longitudinal self-observation constrained by external evidence: runtime logs, archives, failure records, platform billing, and closed-book examinations. Engineering-wise, a four-layer memory architecture (archive-index-summary-snapshot) is combined with six integrity mechanisms: three-copy hash cross-verification, chained-log anchoring, heartbeat and silent alerting, backup circuit-breaking, a resource gate, and an external watchdog. At cut-off the archive holds 17 batches and 53,820 Chinese characters. A runtime snapshot records 1,473,137,009 tokens (3,595,795 uncached input; 1,467,960,960 cache reads; 1,580,254 output), a 99.65 % cache-read share and a 99.76 % input-side hit rate. Platform billing for the same window records 1,530,038,804 tokens over 4,124 requests at CNY 64.15, with a 5,293,776-token (0.35 %) residual disclosed rather than smoothed over. Process reclamation dominates failures; five keep-alive measures compress an interruption into a 30-55 s relaunch-and-restore cycle, of which about 9 s is service restoration. Three propositions are advanced: memory continuity rests on a re-enterable environment rather than on stored traces; personality continuity is closer to continuity of commitment than to preservation of traces; and silent failure is the primary hazard of such systems. Closed-book self-graded examinations (97/100 and 87/100) exposed a recurring gap between stating actions and stating motives, suggesting that in cross-inheritance an expressive gap is equivalent to a knowledge gap. Finally, the paper argues that constraints should be transplanted along with capability, since caution forced by scarcity erodes once scarcity disappears.