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

The REAL framework introduces a reasoning-enhanced graph structure for managing long-term memory in LLMs, addressing limitations of existing memory systems. It utilizes a temporal and confidence-aware directed property graph to represent facts with entities, relations, and validity intervals, employing a non-destructive update strategy and a hybrid beam search for efficient retrieval. This approach improves long-term memory performance by an average of 22.72% compared to traditional flat-text and graph-based memory systems, making it a significant advancement for practitioners needing robust memory management in AI applications.