Rather, the proposal is focusing on what users of these data structures would / could see. The idea is that relational structures have the same ease of use and flexability that things like hashes or arrays or sequences or sets do now. They can of course just be stored in RAM like the aforementioned, when the working set of data is appropriately small, but just as a hash-doing class can have a disk-tied implementation as well, for scalability and/or persistence so can a relation-doing class. And this is one main reason why Relation etc is a role rather than a class, so people can choose how it works.
OK. I can see that a tied-relation could help solve the talking to disk problem. But I wonder about some of the other RBMS things on offer - locking, indices etc? Some of these features are there to assist with getting data efficiently to and from the disk. Although they are not artefacts of the relational model they are important parts of what makes a database work. Could your relational model be "tied" to an existing physical database?