Tag
S3Mem proposes a structured spatiotemporal scene-event memory framework for long-horizon interactive question answering, using anchor-sensitive retrieval and token-budget-aware evidence interface to outperform standard RAG in multiple environments.