VL-MemKnG boosts navigation QA accuracy by 9% with hybrid memory

Answering navigation-related questions from long egocentric videos requires retrieving and organizing evidence scattered across time, a challenge that pure vision-language models struggle with due to high computational cost and inefficiency for repeated queries. VL-MemKnG addresses this by extending the earlier VL-KnG framework with a hybrid memory that combines a structured spatio-temporal knowledge graph (capturing object associations and relational data) with persistent segment-level contextual memory that preserves broader temporal continuity. A joint retrieval-and-reasoning module operates over both representations to produce evidence-grounded answers and temporally organized support.

On the newly introduced WalkieKnowledgeT+ benchmark, designed for long-horizon navigation-oriented video QA with temporally distributed reasoning tasks, VL-MemKnG achieves Top-1 retrieval accuracy of 67% (up from 58%) and Recall@1 of 40.55% (up from 34.50%), outperforming Gemini 2.5 Pro and Qwen 3.5+. The improvements are most pronounced on temporal-global and scattered aggregation questions, demonstrating the value of combining relational memory with segment-level context. This work shows a practical path toward efficient, high-accuracy AI for robotics and autonomous navigation without relying on expensive long-context models.