Chronofy Enhances RAG Systems with Temporal Validity.

Retrieval-Augmented Generation (RAG) systems are powerful for grounding large language model (LLM) outputs with external knowledge. However, a significant challenge is "temporal hallucination," where RAG systems treat all retrieved facts as equally valid, regardless of their age, leading to the use of obsolete information. A new framework, Chronofy, addresses this by embedding temporal validity directly into RAG's core processes.Chronofy introduces a three-layer neuro-symbolic architecture called Temporal-Logical Decay Architecture (TLDA). The first layer structurally integrates fact age into Matryoshka embeddings by reserving a dedicated temporal subspace. This ensures that temporal provenance is an inherent part of the data representation.The second layer enhances retrieval by incorporating learnable exponential decay functions into graph-based retrieval mechanisms, where the decay coefficient is grounded in Bayesian decision theory. Finally, the third layer applies Signal Temporal Logic (STL) robustness functions to evaluate the temporal validity of retrieved knowledge, not just LLM output confidence. It enforces a "weakest-link" principle, bounding output confidence by the most decayed evidence, and can trigger data re-acquisition when temporal context is insufficient. Evaluations on temporal knowledge graph forecasting and QA benchmarks demonstrate Chronofy's effectiveness in improving retrieval precision and reducing temporal hallucination.