HAGE: Harnessing Agentic Memory via RL-Driven Weighted Graph Evolution

Researchers propose HAGE, a weighted multi-relational memory framework that improves how large language model agents retrieve and traverse information by treating memory as a dynamic graph rather than static lookups. The system uses reinforcement learning to optimize edge representations and routing behavior, achieving better long-horizon reasoning accuracy with improved efficiency compared to existing agentic memory systems.

HAGE addresses a fundamental limitation in current LLM-based agent architectures: the assumption that memory retrieval is a static lookup problem. Traditional systems rely on flat vector search or rigid graph structures that cannot capture nuanced relationships between events or adapt to different query contexts. This paper reconceptualizes memory retrieval as a learned, query-dependent traversal problem where relationships themselves become trainable parameters.

The technical innovation centers on relation-specific graph views with trainable edge embeddings that encode multiple relational signals. When an agent queries memory, an LLM classifier identifies the relational intent, and a routing network dynamically adjusts how edges are weighted based on context. This enables the system to suppress noisy connections while emphasizing high-utility paths—mimicking how human memory associatively prioritizes relevant information.

The reinforcement learning framework jointly optimizes routing behavior and edge representations using downstream task performance as the training signal. This end-to-end approach ensures that memory traversal improvements directly translate to better agent reasoning outcomes rather than optimizing proxies. Empirical results demonstrate measurable gains in long-horizon reasoning accuracy alongside favorable efficiency trade-offs.

For the AI development community, HAGE represents progress toward more capable agentic systems that can maintain and leverage complex relational context over extended interactions. The work has implications for autonomous agents operating in knowledge-intensive domains where relationship strength and relevance vary by task. This advancement contributes to building agents with more sophisticated memory management—a critical bottleneck as applications scale to longer planning horizons.

• →HAGE treats memory retrieval as learned, query-conditioned graph traversal rather than static lookup.
• →Trainable edge embeddings encode multiple relational signals with context-dependent weighting.
• →RL-based training jointly optimizes routing behavior and edge representations using downstream task performance.
• →System achieves improved long-horizon reasoning accuracy with better accuracy-efficiency trade-offs than existing methods.
• →Framework enables agents to dynamically suppress noisy relationships while emphasizing high-utility memory paths.