(Follow up to a discussion in the SDL forum.)
From a posting by CafeTran's developer at the CafeTran Freshdesk:
The function has been present in CT for a few years, and if well tuned, it can be used with excellent accuracy both with Auto-assembling and Auto-completion. In particular, see the "Subsegment to Auto threshold" (set to 7 hits by default) option in Edit > Options > Memory tab. If you find that any false positive hits get in the way with Auto-assembling, just increase that number. You can also lower the number to get more hits if your TM is of a good quality (a specialized one). It really boils down to the size and quality of your translation memories to produce satisfactory results with this function.