Emotional Support Conversation (ESC) systems help seekers understand and cope with distress through multi-turn interaction. Existing ESC systems typically improve responses by adding auxiliary support knowledge or prompting LLMs to reason over dialogue context. However, they usually pass inferred seeker information to the generator as a single state, so factual circumstances and mental state inferences are stored, selected, and updated in the same way. This blurs the difference between persistent facts and transient personal states, making it difficult to retain stable grounding while adapting to psychological changes across turns. Therefore, in this paper, we propose MIA, a Personal-Factual Adaptive Memory framework for multi-turn emotional support conversation. MIA first infers Emotional Theory-of-Mind factors and separates them into factual and personal states. It then uses a Personal-Factual Discriminator to decide which state should guide the next response, and an Obsolescence Removal Mechanism to remove outdated memory before generation. This design enables the generator to preserve factual grounding, focus on the turn-relevant support orientation, and avoid stale psychological assumptions. Experiments on ESConv and CPsyCounD show that MIA better understands the seeker’s situation, provides more constructive guidance, and maintains more coherent support across turns. Our codes are available at: https://github.com/ Wh2018/MIA.

Proactive streaming video understanding requires Video-LLMs to decide when to respond as a video unfolds, a task where existing methods often fall short due to their implicit, query-agnostic modeling of visual evidence. We introduce Response-G1, a novel framework that establishes explicit, structured alignment between the accumulated video evidence and the query’s expected response conditions via scene graphs. The framework operates in three fine-tuning-free stages: (1) online query-guided scene graph generation from streaming clips; (2) memory-based retrieval of the most semantically relevant historical scene graphs; and (3) retrieval-augmented trigger prompting for per-frame “silence/response” decisions. By grounding both evidence and conditions in a shared graph representation, Response-G1 achieves more interpretable and accurate response timing decisions. Experimental results on established benchmarks demonstrate the superiority of our method in both proactive and reactive tasks, validating the advantage of explicit scene graph modeling and retrieval in streaming video understanding.