I model the dynamics of the amount of knowledge being learnt and created in the co-invention process between two inventors, by highlighting interpersonal knowledge similarity. Further, I employ U.S. inventor-dyad-level patent data (including 250,546 inventor-dyads) to examine the relationship between knowledge similarity and co-invention performance of the two-inventor teams. The preliminary findings are consistent with the model that predicts a left-skewed inverted U-shaped relationship, suggesting that co-invention performance measured by number of patents is better in teams that consist of inventors with a medium level of similarity in technological knowledge prior to the co-invention process, and that the performance measured by weighted patents is better in teams with more technologically dissimilar members. The study enriches the literature on economics of innovation by explicating knowledge production of inventors in the co-invention process and revealing the left-skewed inverted U-shaped relationship between knowledge similarity and knowledge production performance. In practices, the study suggests assembling teams with technologically dissimilar members for optimal knowledge production performance, especially for the purpose of producing high-quality inventions.