This paper describes user modelling in I-Help, a system to facilitate communication amongst learners. There are two I-Help components: Private and Public Discussions. In the Private Discussions learners take part in a one-on-one interaction with a partner (possibly a peer). The Public Discussions are open - everyone in the group has access to all discussion forums relevant to that group. The Public Discussions are most suited to discussion of issues where there might be a variety of valid viewpoints, or different solutions to a problem. It is also useful for straightforward questions and answers that have wide-spread applicability. The Private Discussions are better suited for more intensive interactions involving peer tutoring or in-depth discussions. Because there is only one helper in such situations, I-Help requires a method of selecting an appropriate helper for an individual. We describe the user modelling that takes place in each part of I-Help, in particular to effect this matchmaking for Private Discussions. This modelling takes advantage of a distributed multi-agent architecture, allowing currently relevant user model fragments in various locations to be integrated and computed at the time they are required.