A key property of natural language is its ability to establish relationships between non-adjacent items in a sentence. For example, a verb must agree with its subject, though the two items may be separated by several words or phrases, as in The student(s) in the IGERT program eat(s) lunch every Thursday. Constructing these linguistic relations requires rapid access to the products of past analyses to integrate incoming material into a developing representation of the sentence. The success of interpretation will then depend on the ability of the parser to retrieve the correct information from memory. Recent research, however, has shown a selective profile for retrieval success in online dependency formation: some dependencies, like subject-verb agreement, are prone to errors, while other dependencies, such as those involving reflexives, are not . In this talk, I will explore why these dependencies show different online profiles with respect to memory access to better understand how memory functions and fails when it is recruited for language processing. To this end, I will focus on a particular dependency that involves null subjects in English: "adjunct control". Adjunct control is a useful test case because it shares properties with both agreement and reflexive licensing. Results from several behavioral studies indicate that susceptibility to errors in memory retrieval is a consequence of the use of specific content cues in retrieval processes.