The current (Jan 2014) Jena in-memory dataset uses a general purpose container that works for any storage technology for graphs together with in-memory graphs.
This project would develop a new implementation design specifically for RDF datasets (triples and quads) and efficient SPARQL execution, for example, using multi-core parallel operations and/or multi-version concurrent datastructures to maximise true parallel operation.
This is a system project suitable for someone interested in datatbase implementation, datastructure design and implementation, operating systems or distributed systems.
Note that TDB can operate in-memory using a simulated disk with copy-in/copy-out semantics for disk-level operations. It is for faithful testing TDB infrastructure and is not designed performance, general in-memory use or use at scale. While lesson may be learnt from that system, TDB in-memory is not the answer here.