HMA-RAG: A Hierarchical Memory Augmented Retrieval Augmented Generation Framework for Fake News Detection

Existing methods for detecting fake news, such as traditional machine learning and transformer-based models, often work as separate classifiers that do not use outside evidence, historical knowledge, or ways to make the results easier to understand. This paper presents a framework that combines retrieval augmentation and hierarchical memory to overcome three primary constraints: Insufficient external evidence, the absence of memory, and restricted explainability in existing methodologies. We present HMA-RAG (Hierarchical Memory-Augmented Retrieval-Augmented Generation), a dual-branch architecture that integrates DistilBERT-based classification with dense retrieval via MiniLM. The framework has a confidence-gated mechanism that only turns on a multi-stage verification pipeline when certain conditions are met. This pipeline includes evidence agreement, memory consistency, and temporal tracking. A single credibility score combines these signals, and a feedback loop keeps long-term memory up to date. Tests on the WELFake dataset (72,134 articles) show that HMA-RAG achieves an F1-score of 98.83%, which exceeds classical baselines and is comparable to a larger transformer model, FakeBERT (98.90% on the ISOT dataset), while using fewer classification parameters (66M vs. approximately 115M). Ablation studies verify the efficacy of the hierarchical memory components. Among the systems compared, only HMA-RAG provides four-layer explainability: Evidence retrieval, dual-stage verification signals, temporal context tracking, and credibility score decomposition. These results indicate that combining retrieval augmentation with hierarchical memory can achieve accuracy comparable to substantially larger models while providing structured, multi-layer transparency intended for deployment in journalism, fact-checking, and content moderation.