Recorded presentations have become pervasive in education, yet they often reduce knowledge transfer to a one-way broadcast: viewers cannot clarify their confusion or gauge whether others share their struggles. We present EchoMind, a system that augments recorded presentations with a conversational agent grounded in presentation materials and collective viewer signals (behavioral traces and affective reactions), alongside timeline-aligned community awareness visualizations. In a between-subjects study (N=75), participants watched a 5-minute recorded lecture under one of three conditions: video-only, video with a generic LLM chatbot, and video with EchoMind. EchoMind significantly improved the subjective learning experience across multiple dimensions, including confusion clarification, in-the-moment expression, and co-learning perception, without increasing perceived workload. The generic LLM condition did not significantly outperform video-only and received the lowest mean ratings on usability. These findings suggest that contextual grounding and community awareness, rather than the presence of conversational AI alone, make AI a meaningful support in asynchronous video learning.