Trust has been established to be a key factor in fostering human-robot interactions. However, trust can change overtime according to different factors, including a breach of trust due to a robot’s error. In this exploratory study, we observed people’s interactions with a companion robot in a real house, adapted for human-robot interaction experimentation, over three weeks. The interactions happened in six scenarios in which a robot performed different tasks under two different conditions. Each condition included fourteen tasks performed by the robot, either correctly, or with errors with severe consequences on the first or last day of interaction. At the end of each experimental condition, participants were presented with an emergency scenario to evaluate their trust in the robot. We evaluated participants’ trust in the robot by observing their decision to trust the robot during the emergency scenario, and by collecting their views through questionnaires. We concluded that there is a correlation between the timing of an error with severe consequences performed by the robot and the corresponding loss of trust of the human in the robot. In particular, people’s trust is subjected to the initial mental formation.