Two experiments were carried out to see what parts of a sentence are best retained, whether those parts retained best immediately after acquisition are also remembered best after a 48 hour retention interval, and whether or not the parts of a sentence which are retained depend on the semantic and syntactic structure of that sentence. In the first experiment, retention was assessed by a recognition method, and in the second experiment, it was measured by cued or primed recall. In the first experiment, high predictability and simple sentences resulted in significantly better recognition of old sentences at both the zero and 48 hour retention intervals. Transformational complexity was not significant, active and passive sentences being recognized equally well. The detection of word changes did not change over the 48 hour retention interval. In the cued recall experiment, high predictability and simple sentences again resulted in significantly better recognition at both time intervals. Another result which paralleled one from the first experiment was the relatively high level of recall of simple unpredictable sentences compared with complex unpredictable ones. The relationship of the findings to various theories of memory for sentences is discussed. (MKM)