Hybrid Graph-Text Retrieval for Multi-Hop Reasoning
Combine graph-based structural navigation with dense text chunk properties in graph nodes to improve multi-hop reasoning performance.
Source paper
Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, Deshraj Yadav
Generated by AI. The paper’s authors did not write these.
Combine graph-based structural navigation with dense text chunk properties in graph nodes to improve multi-hop reasoning performance.
Implement a dual-stage memory process with an asynchronous consolidation agent to handle fact merging and conflict resolution without synchronous latency.
Replace memory deletion with temporal versioning using valid time stamps to preserve historical fact timelines in user profiling systems.
This paper introduces Mem0 and Mem0-graph, memory architectures designed to enable Long Large Language Models (LLMs) to maintain coherence across extended, multi-session interactions. By dynamically extracting salient facts into a vector store or a knowledge graph and employing an incremental update mechanism (add/update/delete), the system achieves state-of-the-art performance on the LOCOMO benchmark while significantly reducing latency and token costs compared to full-context or RAG baselines.
The system uses a pipeline where an LLM analyzes message pairs to extract facts, which are then managed via tool-calls (ADD, UPDATE, DELETE) in a vector database (Mem0) or a Neo4j knowledge graph (Mem0-graph). Retrieval is performed using dense vector similarity for text or a dual entity/triplet matching strategy for graphs.