Behavioral studies have shown that verbal information is better retained when it is self-generated rather than read (learned passively). We used fMRI and a paired associates task to examine brain networks underlying self-generated memory encoding. Subjects were 49 healthy English speakers ages 19-62 (30 female). In the fMRI task, related word pairs were presented in a "read" condition, where subjects viewed both words and read the second word aloud, or a "generate" condition, where the second word was presented with only the first letter and the subject was required to generate the word. Thirty word pairs were presented in each condition. After the fMRI scan, words that were read or generated were presented, each with two foils, in a forced-choice recognition task. On the recognition post-test, words from the "generate" condition were more correctly recognized than from the "read" condition (80.0% for generated words versus 72.0% for read words; t(48)=5.17, p<0.001). FMRI revealed increased activation for generate>read in inferior/middle frontal gyri bilaterally (L>R), anterior cingulate, and caudate nucleus and the temporo-parietal-occipital junction bilaterally. For the "read" condition, better subsequent memory performance across individual subjects was positively correlated with activation in the cuneus bilaterally. In the "generate" condition, better subsequent memory performance was positively correlated with activation in the left superior temporal gyrus. These results suggest that self-generation improves memory performance, that enhanced cortical activation accompanies self-generated encoding, and that recruitment of a specific brain network underlies self-generated encoding. The findings may have implications for the development of procedures to enhance memory performance.