Enhancing continuous queries over data streams with temporal functions and predicates enriches the expressive power of those queries. While traditional continuous queries retrieve only the values of attributes, temporal continuous queries retrieve the valid time intervals of those values as well. Correctly evaluating such queries requires the coalescing of adjacent timestamps for value-equivalent tuples prior to evaluating temporal functions and predicates. For many stream applications, the available computing resources may be too limited to produce exact query results. These limitations are commonly addressed through load shedding and produce approximated query results. There have been many load shedding mechanisms proposed so far, but for temporal continuous queries, the presence of coalescing makes theses existing methods unsuitable. In this paper, we propose a new accuracy metric and load shedding algorithm that are suitable for temporal query processing when memory is insufficient. The accuracy metric uses a combination of the Jaccard coefficient to measure the accuracy of attribute values and $mathcal{PQI}$ interval orders to measure the accuracy of the valid time intervals in the approximate query result. The algorithm employs a greedy strategy combining two objectives reflecting the two accuracy metrics (i.e., value and interval). In the performance study, the proposed greedy algorithm outperforms a conventional random load shedding algorithm by up to an order of magnitude in its achieved accuracy.