The Company Line: How Shared Organizational Context Can Narrow Advice and Cause Hidden Harm

Memory systems are increasingly being used to manage context and interaction history better, both to retain personal history and to share information across organizations. In this work, we study the effect such memory systems may have on the \emph{diversity} of model outputs in a ranked recommendation setting. Specifically, we simulate a team of professionals with distinct personas in domains including financial advising, data analysis and movie recommendation. Each professional has an AI advisor that utilizes personal memory or shares organizational context for making recommendations about assigned work. We observe that increased personal context leads to lower intra-user diversity with drops of up to 14\%. In contrast, the context shared between users tends to decrease inter-user diversity with drops of 13--20\%. Across these context management policies, while average accuracy of recommendations remains similar, accuracy for certain subgroups drops significantly. Thus, both the diversity and subgroup-level utility may be negatively impacted by some policies of context management within an organization.