The self-teaching hypothesis posits that enduring orthographic and phonological representations are produced when children independently recode print into speech. However, very little research has examined how children self-teach when initial decoding attempts are weak or ineffective. In this within-participant design, 25 students in Grade 2 learned to read 85 different words in 4 conditions. Words were read in and out of context, with and without feedback. Accuracy rates were recorded throughout 5 training sessions (2 word repetitions per session = 10 repetitions in total). A posttest was administered after a 6-day delay by reinstating the training materials. At the end of training, the highest accuracy scores were observed when children read in context/feedback followed by when they read in isolation/feedback, and then in context/no feedback; the lowest accuracy scores were observed when children read in isolation/no feedback. This pattern remained over the retention period, suggesting that external support from feedback, and top-down support from context, can help create word representations in memory. The results are discussed in relation to the importance of whole-word phonology within self-teaching.