Volatility-Driven Decay: Adaptive Memory Retention for RAG Systems Under Unknown Drift

VDD is a control-theory approach for adaptive memory retention in RAG systems. The decay rate is dynamically modulated by environmental volatility instead of being fixed, achieving ‘never worst’ performance across drift patterns without manual tuning. Validated on 120 real-world facts across 3 domains (React, Python, Node.js) in 42 experiments. Abstract Retrieval-Augmented Generation (RAG) systems face a silent reliability crisis: when knowledge drifts—APIs deprecated, guidelines revised, frameworks migrated—retrieval continues serving outdated answers with full confidence. We find that naive retrieval serves stale content 67% of the time across 120 real-world technical facts. Existing solutions assume known drift patterns: recency-biased retrieval suits constant change, time-weighted decay handles gradual transitions, and sliding windows work for bounded contexts. None adapts when the pattern is unknown—the typical production condition. We introduce Volatility-Driven Decay (VDD), where the memory decay rate λ(t) dynamically adjusts based on detected semantic drift via sigmoid activation on embedding centroid distance. Structurally inspired by control-theoretic adaptive filtering and cognitive models of memory consolidation, VDD continuously modulates its forgetting rate between a conservative baseline and an aggressive ceiling. Validated on 120 facts across React, Python, and Node.js documentation (n=30, bootstrap CIs, Benjamini-Hochberg corrected, Cohen’s d), VDD achieves 87.5% error reduction versus static baselines, ranks #2 among 13 methods across four drift patterns, and is never the worst performer in any tested scenario. An auto-calibration method (75th percentile of burn-in volatilities) eliminates manual V0 tuning, improving over hand-tuned defaults by 22.7% on average. We are equally explicit about limitations: recency outperforms VDD in permanent high-drift (d = −2.90), time-weighted retrieval dominates gradual transitions (0.871 vs 0.584), and adaptive baselines (Holt-Winters, EMA-λ) beat VDD in 3/4 synthetic scenarios. On StreamingQA (36K temporal QA pairs), all decay methods including VDD underperform no-decay (d = −7.68), confirming that VDD is counterproductive on knowledge-accumulation tasks—it is specifically designed for knowledge-replacement scenarios. On FreshQA (452 temporal questions), all decay methods achieve 100% current-answer retrieval rate versus no-decay’s 70–80%, though keyword-overlap scoring fails to capture this advantage. Rankings transfer between hash-based and real neural embeddings (4/5 scenario agreement, r = 0.935 effect size correlation) and across embedding models (nomic-embed-text vs mxbai-embed-large, Spearman ρ = 0.978), though effect sizes attenuate with real embeddings (mean ratio 0.70x), indicating controlled-condition results are directionally valid but inflated. VDD’s value is precisely for production systems where drift patterns cannot be predicted in advance. This positioning complements—rather than competes with—modern RAG architectures (Self-RAG, CRAG, Adaptive-RAG) that improve what to retrieve, while VDD addresses when to forget. We release the complete benchmark, dataset, and evaluation framework. Opinions are my own. This work does not relate to my position at Amazon.