ForgetMeNot: What and how users expect intelligent virtual agents to recall and forget personal conversational content

An important aspect of human interaction is our ability to store, retain, recall and organise information. Memory assists in building rapport and gaining trust. The growing field of artificial companions involving long term and casual/social relationships will require appropriate handling of memories of the user by an Intelligent Virtual Agent (IVA). This is a first study to determine what types of personal information (i.e.; domain/task and casual/social) users expected an IVA to discuss and remember and how users responded to various levels of IVA recall and forgetting of that information: complete recall; total loss of recall; partial recall and incorrect recall. Our experiment collected conversations and survey responses from participants who interacted with a virtual real estate agent according to one of two treatments over three experimental sessions. No significant differences were found in participants' responses between task-focused or casual/social content, however a larger sample size may produce different results and some participants were not comfortable with answering the social questions. Recall was found to increase the user's enjoyment of interacting with the agent and also to increase the believability of the character over multiple interactions. We found that characters who recall information incorrectly are highly frustrating, are seen as having unnatural memory and this decreases the believability of the character. However, characters that exhibit forgetting, either explicitly stating their forgetfulness or not mentioning it at all, are seen to have a more natural memory and can help to increase the believability of the character. The study also suggests that forgetting affects the level of trust the user feels for the character.