Standard sliding context windows in AI agent systems evict data based solely on recency, frequently discarding critical foundational information after periods of inactivity. To address this structural limitation, developer Emmimal has introduced a usage-reinforced memory decay engine that applies the Ebbinghaus forgetting curve to agent working memory. The system replaces rigid recency cutoffs with a dynamic retention model where repeated recall events compound an item stability, significantly extending its operational horizon. The engine operates on a deterministic turn counter rather than wall-clock time. Each memory item is assigned a base stability value, and its retention score decays exponentially over elapsed turns. Every successful recall non-linearly increases the item stability, effectively flattening the decay curve. This architecture ensures that frequently referenced rules, constraints, or facts remain accessible regardless of intervening conversation noise. In rigorous benchmarking across fifty seeded sessions of 150 turns each, the decay engine achieved a 100 percent foundational recall rate, compared to a zero percent retention rate for standard recency-only baselines. Both systems demonstrated comparable token-footprint reduction, pruning roughly 90 percent of peripheral noise. The performance gap emerged exclusively during silent intervals following early reinforcement, confirming that recency windows fail when importance and timing diverge. The development process revealed two critical implementation pitfalls. Initial testing used a stability value too low, causing unreforeced items to expire before their first scheduled recall, rendering the system ineffective. A second benchmarking flaw involved uniformly random recall scheduling, which artificially guaranteed baseline failure by pushing recalls past expiration windows. Correcting the schedule to mimic spaced-repetition intervals and adjusting baseline stability to eight restored valid comparative metrics. All results were verified as byte-identical across Linux and Windows environments, confirming full determinism and reproducibility. Despite its effectiveness, the system operates within defined boundaries. It only preserves information that receives repeated use; single-exposure facts decay at standard rates. The benchmark approximates token reduction by item count rather than actual character length, and real-world deployment requires explicit recall detection logic to differentiate contextual presence from functional utility. Additionally, eviction thresholds and stability parameters must be calibrated to specific session rhythms, as burst-and-silence patterns do not universally apply. The findings underscore a fundamental shift in AI memory architecture: retention should be measured by functional dependency rather than temporal proximity. By decoupling memory survival from arbitrary recency limits, the Ebbinghaus-based engine provides a deterministic, low-overhead alternative for long-horizon agent tasks. The complete implementation, benchmarks, and failure-case analysis are publicly available for independent verification and production adaptation.