In this extended abstract, we investigate the design of learning representation for human intention inference. In our designed human intention prediction task, we propose a history encoding representation that is both interpretable and effective for prediction. Through extensive experiments, we show our prediction framework with a history encoding representation design is successful on the human intention prediction problem.