HAGE: Harnessing Agentic Memory via RL-Driven Weighted Graph Evolution
Researchers have introduced HAGE, a novel weighted multi-relational memory framework designed to enhance memory retrieval in agentic large language model (LLM) systems. Unlike traditional methods that rely on static vector searches or fixed binary graphs, HAGE treats retrieval as a sequential, query-conditioned traversal over a unified relational memory graph. The system organizes memory into relation-specific graph views, where edges carry trainable feature vectors encoding various relational signals. An LLM-based classifier identifies relational intent from queries, allowing a routing network to dynamically adjust edge embeddings. This mechanism prioritizes high-utility paths while suppressing noise. Furthermore, HAGE employs a reinforcement learning-based training framework to jointly optimize routing behavior and edge representations using downstream tasks. Empirical results indicate that HAGE significantly improves long-horizon reasoning accuracy and offers a better accuracy-efficiency trade-off compared to current state-of-the-art agentic memory systems. The source code for this innovation is publicly available, marking a significant step forward in dynamic memory management for AI agents.
Wire timeline
HAGE: Harnessing Agentic Memory via RL-Driven Weighted Graph Evolution
Researchers have introduced HAGE, a novel weighted multi-relational memory framework designed to enhance memory retrieval in agentic large language model (LLM) systems. Unlike traditional methods that rely on static vector searches or fixed binary graphs, HAGE treats retrieval as a sequential, query-conditioned traversal over a unified relational memory graph. The system organizes memory into relation-specific graph views, where edges carry trainable feature vectors encoding various relational signals. An LLM-based classifier identifies relational intent from queries, allowing a routing network to dynamically adjust edge embeddings. This mechanism prioritizes high-utility paths while suppressing noise. Furthermore, HAGE employs a reinforcement learning-based training framework to jointly optimize routing behavior and edge representations using downstream tasks. Empirical results indicate that HAGE significantly improves long-horizon reasoning accuracy and offers a better accuracy-efficiency trade-off compared to current state-of-the-art agentic memory systems. The source code for this innovation is publicly available, marking a significant step forward in dynamic memory management for AI agents.
cs.AI updates on arXiv.org