Title: MemORAI: Enhancing LLM Conversational Agents through Adaptive Graph Intelligence for Memory Organization and Retrieval

Abstract: Current Large Language Models (LLMs) struggle to maintain persistent memory for long-term, personalized interactions. Existing graph-based memory architectures are often hindered by issues such as information dilution, a lack of provenance tracking, and rigid retrieval mechanisms that fail to account for query context. To address these challenges, we propose MemORAI (Memory Organization and Retrieval via Adaptive Graph Intelligence), a novel framework incorporating three key innovations. First, it employs selective memory filtering combined with dual-layer compression to preserve content relevant to the user persona. Second, it utilizes a provenance-enriched multi-relational graph to track the origins of facts at the individual turn level. Third, it implements query-adaptive subgraph retrieval using Dynamic Weighted PageRank, which applies edge weights conditioned on the specific query. Our evaluation on the LOCOMO and LongMemEval benchmarks reveals that MemORAI delivers state-of-the-art results in both memory retrieval and personalized response generation. These findings underscore the critical importance of selective storage, enriched representation, and adaptive retrieval in enabling coherent and personalized LLM agents.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC