When referring to named objects, speakers can choose either a name (mbira) or a description (that gourd-like instrument with metal strips); whether the name provides useful information depends on whether the speaker’s knowledge of the name is shared with the addressee. But, how do speakers determine what is shared? In 2 experiments a naïve participant (director) learned names for novel objects, then instructed another participant (matcher), who viewed 3 objects, to click on the target object. Directors learned novel names in 2 phases. First, the director and the matcher learned (shared) names either together or alone; second, the director learned (privileged) names alone. Directors typically used a name for items with shared names and a description for items with privileged names. When the director and matcher learned the names individually but with knowledge of what the other learned, directors were much more likely to use privileged names than when director and matcher learned shared names together. Experiment 1b separated effects of collaborative learning from partner-specific effects, showing collaborative learning experience with 1 person helps a speaker distinguish shared and privileged information with a new partner who has the same knowledge. Experiment 2 showed that partner-specific effects persisted even when semantic category was a reliable cue to which names were privileged. The results are interpreted as evidence that ordinary memory processes provide access to shared knowledge in real-time production of referring expressions and that shared experience when learning shared names provides a strong memory cue to the ground status of names.