REAL: A Reasoning-Enhanced Graph Framework for Long-Term Memory Management of LLMs

A new framework called REAL proposes managing long-term memory for large language models by structuring conversational history as a temporal, confidence-aware directed property graph, according to a paper submitted to arXiv on 9 Jun 2026 [1]. Large language models, or LLMs, operate within a finite context window, which prevents them from retaining all past interactions. This constraint makes external long-term memory systems necessary for storing, updating, and retrieving historical information [1]. The paper’s authors argue that existing memory approaches suffer from three core weaknesses: flat text-based organizations that miss explicit relationships between memories, structured systems that destructively overwrite evolving facts, and retrieval mechanisms that remain passive when evidence is incomplete [1]. REAL addresses these gaps by representing each atomic fact with entities, relations, valid-time intervals, confidence scores, and exploration intent labels [1]. During memory construction, the framework uses a non-destructive temporal update strategy that preserves parallel fact versions and their validity intervals, allowing it to track how facts change over time [1]. When retrieving information, REAL anchors query-relevant root entities, decouples their exploration intents, and performs a semantic evaluator-guided hybrid beam search to extract compact memory subgraphs [1]. The framework also incorporates counterfactual inference to repair unreliable retrieval states and recover missing memory evidence through implicit logical relations [1]. Comprehensive experiments reported in the paper show that REAL achieves an average improvement of 22.72% over flat-text, graph-based, and other existing memory baselines [1]. The paper was posted on arXiv, an open-access repository for electronic preprints that, as of November 2024, receives about 24,000 submissions per month [11]. arXiv hosts papers across fields including computer science, physics, and mathematics, and has served as a primary distribution channel for machine-learning research during the AI boom that accelerated after the introduction of the transformer architecture in 2017 [3][11]. The REAL framework builds on a broader lineage of work in machine learning, a subfield of artificial intelligence concerned with statistical algorithms that learn from data and generalize to unseen tasks [6]. Modern LLMs themselves rely on neural-network architectures, particularly transformers, which use attention mechanisms to model long-range dependencies in data [7].