Human knowledge consists of schemata based on the specialized procedures used to interpret events in the environment. New schemata are created by modifying models of existing schemata, i.e., learning by analogy. This means that a new situation will be interpreted in accordance with the schema most similar to it in the learner's repertoire. If the schema is an inappropriate fit, a new one based on the original model will be created. By carrying over existing features of a current schema, we can make inferences about a new situation without explicit knowledge of that situation. If the analogy is a good one, most of the inferences from the original schema or source domain will be appropriate to the new one. Conceputal models should have the following properties: (1) it should be based on a domain about which the student is knowledgeable, (2) the target domain and source domain should differ by a minimum number of dimensions, (3) operations natural within the target domain should also be natural in the source domain, and (4) operations inappropriate within the target domain should also be inappropriate within the source domain. Examples from several domains include turtle geometry, kinship terms, and learning a computer text editor. (Author/BK)