I had a post up yesterday over at Plum Line talking about the myth of momentum, and why we shouldn't expect change in one direction in presidential general elections to build on itself. You should also read John Sides, who talks about media coverage, and Nate Silver, who supplies a useful definition and a bit of empirical evidence. What I add to the mix in my post is why we shouldn't expect the groups who are responsible for changes in polls to yield continued changes.
Just one quick addition to all of that: just because momentum is unlikely in the current election doesn't mean it can never happen. Kevin Collins points out that momentum makes sense in multicandidate contests when strategic voting comes into play. That's true.
Momentum also may be relevant in primary elections. Not just because of strategic voting (learning that my favorite candidate has little chance may shift my vote to my second-favorite, in order to beat the candidate I really don't like), but also because of the information environment and the nature of the electorate. First, we have an electorate of generally unattached voters who could vote for any of a number of candidates. Then, a candidate who wins a primary typically get the lion's share of press coverage immediately after that event. I may, as an unattached voter, gravitate towards that winner just because she's the only candidate I'm really aware of that week. Indeed, that dynamic could easily play out in a single-election primary (such as a Senate primary) with a positive poll, debate performance, or financial report serving as the event which allows one candidate to dominate the information environment. Note, however, if parties are able to dominate the information environment, then you should get a lot less momentum.
The main idea, at any rate, is that for "momentum" to work it requires the first set of changes to then lead to further changes -- which requires some pot of voters who will react either will react, but only more slowly, to the cause that produced the first set of changes, or a pot of voters who will react to the change itself. So find conditions where that may work, and you'll find possibilities for momentum.