ARIA: Autonomous Retrieval and Integration Architecture for Continual Knowledge Updating in Large Language Models

Large language models (LLMs) suffer from a fundamental limitation: their knowledge is frozen at the time of training, becoming stale as the world evolves. We present ARIA (Autonomous Retrieval and Integration Architecture), a production system that eliminates this limitation through three coordinated mechanisms: (1) a Data Translation Layer (DTL) — a nine-stage automated pipeline that continuously harvests, denoises, and structures knowledge from the open web; (2) a dual-memory architecture combining a frozen base LLM with a dynamically updated vector store and temporal knowledge graph; and (3) a weekly LoRA micro-adaptation loop guarded by Elastic Weight Consolidation (EWC) to prevent catastrophic forgetting. ARIA achieves daily knowledge currency at approximately 500/day operating cost — roughly 1, 000x cheaper than equivalent full fine-tuning cycles. On a curated temporal reasoning benchmark (ARIA-TRB), ARIA outperforms a static GPT-4-level model by +34. 2 F1 points on questions about events occurring after the base model's training cutoff, while retaining 98. 7% of pre-update benchmark performance across 47 consecutive weekly update cycles. The full system, training code, Docker deployment, and benchmark dataset are released under the Apache 2. 0 license at https: //github. com/Diwakarsrd/ARIA-System.