Aaah, so, you do not need to separate by time signals, but can use geometry! We use(d) silicon probes with up to 40µm c/c site distance and where able to see one unit on two or three sites (see fig for a 8channel Si probe).... We call that: Searching the footprint.
Check out the work of Tim Blanche (Blanche, T. J., M. A. Spacek, J. F. Hetke and N. V. Swindale (2005). "Polytrodes: High-density silicon electrode arrays for large-scale multiunit recording." J. Neurophysiol. 93: 2987-3000.) or with more mathematical sophistication the work of Karim Oweiss on Wavelet Package usage:
Oweiss, K. G. and D. J. Anderson (2002). "Spike sorting: a novel shift and amplitude invariant technique." Neurocomputing 44-46: 1133-1139.
Eleryan, A., M. Vaidya, J. Southerland, I. S. Badreldin, K. Balasubramanian, A. H. Fagg, N. Hatsopoulos and K. Oweiss (2014). "Tracking single units in chronic, large scale, neural recordings for brain machine interface applications." Front Neuroeng 7: 23.
'Wavemark' in Spike2 might be of help so Cambridge Electronic Design (CED) might be able to advise. I only have the problem of separating relatively few separate spikes in one channel so have no direct experience of your situation, but electrode size and spatial configuration of recording pairs of electrodes would be the way ahead - if practical that is!