SAGE: Self-Evolving Graph Memory Engine for Language Agents
A novel AI framework named SAGE (Self-evolving Agentic Graph-memory Engine) has been unveiled to overcome the limitations of long-term memory in language agents. In contrast to current systems like RAG and GraphRAG, which utilize memory graphs as unchanging retrieval middleware, SAGE conceptualizes graph memory as an evolving long-term memory foundation. It integrates two functions: a memory writer that gradually builds structured graph memory from past interactions and a memory reader based on a Graph Foundation Model that retrieves information and offers feedback to the writer. The framework is supported by comprehensive theoretical analyses. SAGE demonstrates enhancements in evidence recovery, answer grounding, and retrieval performance across various benchmarks, including multi-hop QA and open-domain retrieval. The research paper can be found on arXiv with the identifier 2605.12061.
Key facts
- SAGE stands for Self-evolving Agentic Graph-memory Engine.
- It addresses long-term memory bottlenecks in language agents.
- Existing RAG and GraphRAG systems use static memory graphs.
- SAGE models graph memory as a dynamic long-term memory substrate.
- It couples a memory writer and a Graph Foundation Model-based memory reader.
- The memory writer constructs structured graph memory from interaction histories.
- The memory reader performs retrieval and provides feedback to the writer.
- The framework includes rigorous theoretical analyses.
- SAGE was tested on multi-hop QA, open-domain retrieval, domain-specific review QA, and long-term agent-memory benchmarks.
- It improves evidence recovery, answer grounding, and retrieval performance.
- The paper is on arXiv with identifier 2605.12061.
Entities
Institutions
- arXiv