Standard AI memory systems often discard information based solely on recency, treating crucial foundational facts and trivial logs equally once they exit a "context window." This article introduces a novel "usage-reinforced decay engine" that leverages the Ebbinghaus forgetting curve to enhance AI agent memory.

This new approach reinforces an item's stability with each recall, pushing its eviction horizon out non-linearly. In 50 seeded sessions, the decay engine achieved a 100% Foundational Recall Rate, outperforming a recency-only baseline that retained 0%. The engine's effectiveness stems from its understanding that memory stability increases with usage, a concept rooted in psychology dating back to 1885.

The system is deterministic and relies on an explicit turn counter. It proves most effective in scenarios where critical information is established early, goes dormant, and is needed later, common in tasks like coding or customer support. While not a universal memory upgrade, it addresses a structural limitation in current AI memory architectures, ensuring vital information is retained beyond simple recency.