The ability to generalize from the known to the unknown is crucial for learning, and this ability is known to develop early in life. However, the mechanisms underlying early generalization remain contested. Some researchers argue that even early in development people rely on conceptual assumptions to perform induction. In particular, people, including very young children, assume that animals that share category labels belong to the same kind, and animals of the same kind share many properties. Therefore, according to this approach, people rely on common labels to first categorize presented entities, and then to perform category-based induction. However, other researchers suggest that early induction is similarity-based, and that labels are features of objects contributing to the overall similarity of entities. Furthermore, according to this approach, the ability to perform category-based induction is a product of learning. The current series of experiments was designed to differentiate between these theoretical approaches to early induction. Present work argues that if different processes underlie early and mature induction, young children and adults should form different memory traces when performing induction, and therefore, patterns of memory accuracy can be used to infer mechanisms underlying induction at different points in development. In particular, similarity-based induction should lead to encoding of item-specific information and therefore to accurate recognition on a subsequent memory test. At the same time, category-based induction should lead to encoding of predominantly category-specific information, and therefore to poor recognition. Results of 10 experiments reported below suggest that 5- and 7-year-olds spontaneously perform similarity-based induction when presented with members of familiar categories, whereas adults perform category-based induction. Furthermore, category labels do not automatically promote category-based induction in young children. However, children can be trained to perform category-based induction by relying on shared labels, and retention of training is a function of age: 7-year-olds are more likely to retain training over short delays than 5-year-olds. Moreover, under certain conditions, associative training in label-based induction can be as effective as conceptual training. These results cast doubt on the assertion that young children rely on conceptual assumptions to perform induction, and support the similarity-based approach to early induction.