ReKV introduces a training-free approach that revolutionizes video analysis by enabling real-time question answering on streaming video content.

• Selective memory retention - Stores only relevant video information
• Immediate response capability - Answers questions as soon as relevant content appears
• Integration with existing Video-LLMs - Works without modifying base models
• Efficient processing - Eliminates need to reanalyze entire videos for each new query

Security Impact: ReKV enables real-time video surveillance analysis, allowing security systems to answer questions about video content as events unfold without waiting for complete processing - critical for time-sensitive threat detection and response.

Streaming Video Question-Answering with In-context Video KV-Cache Retrieval