This document proposes a new method called temporal multiactivity graph indexing (tMAGIC) to efficiently detect occurrences of high-level activities from data streams. tMAGIC uses a temporal probabilistic graph structure to model activities and link observations over time. Algorithms are introduced for indexing observations into the tMAGIC structure as they occur in constant time. Two problems are defined - finding all activity occurrences above a probability threshold within a sequence, and identifying the activity that best matches a given sequence. Restrictions and pruning are used to make the inherently exponential problems linear in complexity to the number of observations. Experiments confirm tMAGIC has linear time and space complexity and can efficiently retrieve activity instances.