Not All Memories Age the Same: Autodiscovery of Adaptive Decay in Knowledge Graphs

Current knowledge graphs treat all facts as equally current, applying uniform forgetting curves regardless of knowledge type. Researcher Mandar Karhade shows this is fundamentally flawed: different facts (e.g., a person's birthplace vs. their job title) exhibit different temporal dynamics. He proposes a hierarchical framework that replaces uniform decay with a continuous decay surface parameterized by two orthogonal signals: velocity (how frequently a concept is observed) and volatility (how much the value changes between observations, measured via embedding distance). The decay surface is decomposed into three learnable levels: domain-level parameters capture universal patterns (some predicates are permanent, others transient), context-level parameters capture setting-dependent variation, and entity-level adaptation personalizes decay to specific subjects. All parameters emerge from data through survival analysis on observed value lifetimes, requiring no predefined taxonomies or domain expertise.

Experiments on synthetic temporal knowledge graphs demonstrate recovery of planted hierarchical parameters with HDBSCAN ARI = 1.0. Validation on 107 Wikipedia articles and 1,163 patient records from the Synthea clinical EHR simulator shows that velocity-volatility clusters emerge naturally, align with observable persistence patterns, and near-universally exhibit the Lindy effect (Weibull shape k < 1). Critically, uniform decay performs 18x worse than no temporal weighting at all. Heterogeneous decay recovers from this, with each hierarchy level contributing measurable improvement. The findings have direct implications for retrieval-augmented generation, recommendation systems, and any application where knowledge freshness matters.