HAGE: Weighted Multi-Relational Memory for LLM Agents
A recent publication on arXiv (2605.09942) presents HAGE, a framework designed for agentic large language model (LLM) systems that utilizes a weighted multi-relational memory approach. In contrast to conventional retrieval techniques like flat vector searches or static binary graphs, HAGE structures memory as relation-specific graph representations over common nodes, where each edge is associated with a trainable feature vector that captures various relational signals. When a query is made, a classifier based on LLM determines the relational intent, while a routing network adjusts the dimensions of edge embeddings. The traversal scores integrate semantic similarity with learned relational characteristics, facilitating query-conditioned sequential navigation through a cohesive relational memory graph.
Key facts
- Paper arXiv:2605.09942 proposes HAGE framework
- HAGE stands for Harnessing Agentic Memory via RL-Driven Weighted Graph Evolution
- Memory is organized as relation-specific graph views over shared nodes
- Each edge has a trainable relation feature vector
- An LLM-based classifier identifies relational intent per query
- A routing network dynamically modulates edge embedding dimensions
- Traversal scores combine semantic similarity with learned features
- Framework enables query-conditioned sequential traversal
Entities
Institutions
- arXiv