MERMAIDS: Structured Memory and Evidence Reuse for Reducing Multi-Hop Retrieval Hallucinations in Agentic RAG

Agentic Retrieval-Augmented Generation (RAG) systems have advanced the ability of large language models to handle complex, multi-step questions by dynamically planning and executing retrieval operations. However, these systems remain vulnerable to intermediate reasoning breakdowns, redundant retrieval, and evidence inconsistency, which collectively cause answer drift in multi-hop question answering. We propose MERMAIDS, a framework that augments Agentic RAG with a Structured Evidence Memory (SEM) module and a Cross-task Evidence Reuse Cache (ERC). Extracted evidence is organized into a lightweight knowledge graph following a claim-evidence-source-time schema, and a dedicated Conflict Detection module identifies contradictions and triggers selective re-retrieval or answer confidence degradation. Experiments on HotpotQA, MuSiQue, MultiHop-RAG, and FEVER show that MERMAIDS achieves 5.9–8.6 point improvements in exact match accuracy over strong baselines, while reducing redundant retrieval steps by 28–35% and attaining conflict detection precision above 81%.