PrivGemo: Privacy-Preserving Dual-Tower Graph Retrieval with Memory Augmentation Empowering LLM Reasoning

PrivGemo is a privacy-preserving retrieval-augmented framework that, through dual-tower graph retrieval and memory augmentation, enables LLMs to perform multi-hop, multi-entity reasoning over private knowledge graphs. Its dual-tower design keeps the original graph local, bases remote reasoning on an anonymized view beyond entity-name masking, and reduces remote interactions with a hierarchical controller and privacy-aware experience memory. It achieves overall SOTA on six benchmarks, surpassing the strongest baseline by up to 17.1%, and enables Qwen3-4B to achieve reasoning performance close to GPT-4-Turbo.