Quite often, at conferences or while attending seminars, I will start an interesting discussion with the speaker, first on-site, then later by email, and even though at some point there seems to be some mutual interest, it almost never gives an actual collaboration (i.e. working on an actual paper).
I have no particular problem of working with different people, so I was wondering if it was quite usual to have this huge ratio of "collaboration failure"? In particular, my problem is that, although it's quite simple to have an idea, it seems quite hard to do the next step, that is to actually work with someone you have no connection, and who might even live in a different country. Are there some techniques to make a "temporary" collaboration work, or at least to detect those which are unlikely to work?
In my experience, starting a collaboration is incredibly easy: you use your network of contacts to identify someone who'd be willing and interested in solving a problem. You talk at a conference or meeting, or arrange a visit to their laboratory.
Maintaining a collaboration, however, is next to impossible. It only works if you have a history of successful results early on, or if you have already had a long history of acquaintance with one another before the collaboration began. (In other words, were you friends or colleagues before the work started?)
Otherwise, I would recommend making sure that you start off with "low-hanging fruit": problems that can be solved mutually within the framework of existing funding on both of your parts, with value for both of you. This is important because one of the challenges of getting grants is that reviewers for funding agencies typically want to see an existing record of collaboration—mutual publications and effort—before they're ready to award money to a new collaborative proposal. There are exceptions to this, but they're by no means common.
After that, you have a track record of working together which will let you grow the collaboration into something further.