Profile-Graph Memory for LLM Agents: Implicit Cross-Entity Traversal through Narrative Profiles

Short summary

Researchers introduce MemHop, a 1,000-question multi-hop memory benchmark across 10 social-network scenarios, and ProGraph, a two-layer memory architecture using profile expansion and compression residuals. ProGraph matches FullContext on MemHop (80.1%) and exceeds it on LoCoMo (78.4%), outperforming Mem0, A-Mem, HippoRAG, and RAG. Ablations show profile expansion drives multi-hop reasoning while compression residuals improve precision recall.

• •MemHop: new multi-hop memory benchmark with 1,000 questions at hop depths 1-5
• •ProGraph: two-layer memory combining profile expansion and compression residuals
• •Outperforms Mem0, A-Mem, HippoRAG, and RAG on both benchmarks

Generated with AI, which can make mistakes.