What's a ROC?
Imagine you conduct the following experiment.  Subjects study a list of words.  They then take a recognition test.  The test includes some old words and some new words.  They respond by using the following confidence scale:
-3 -2 -1 +1 +2 +3
Absolutely Certain New Moderately Certain New Somewhat Certain New Somewhat Certain Old Moderately Certain Old Absolutely Certain Old
An ROC curve curve plots the cummulative proportion of old versus new items that are recognized at each level on the confidence scale.  So to create an ROC curve the first step is to determine the proportion of old and new items where the level of confidence is at least -2, at least -1, at least +1, at least +2, and at least +3.  You don't include at least -3 because it would always be the case that 100% of the scores are at least -3.  Here's some data that I made up for illustrative purposes.
Proportion of responses that are...  Old Items New Items
At least -2 0.95 0.85
At least -1 0.87 0.70
At least +1 0.60 0.42
At least +2 0.43 0.10
At least +3 0.25 0.01
 
These scores are then plotted to form the ROC curve:

ROC curves tell you what the hit rate (i.e. correct "old" responses) and false alarm rate (i.e. incorrect "old" responses) is at different levels of confidence.  Consider the point on the lower left of the ROC curve.  That point represents the highest confidence level (+3).  Twenty-five percent of the targets are recognized at that level of confidence while only 1% of the lures are.  Consider the next point.  That point represents the proportion of items recognized with a confidence of at least +2.  Forty three percent of the targets are recognized with a confidence level that is at least +2 while 10% of the lures are.  As you go from the the top right to the bottom left the scores represent progressively more conservative responses.  The scores on the bottom left are only those that subjects are absolutely certain of.

ROC curves are useful because theories of memory often make predictions about what ROC curves should look like.  For example, single process signal detection models (i.e. familiarity only models) predict that ROC curves should look like this:


Notice that this curve intercepts at the points (0,0) and at (1,1).  Yonelinas's dual process model, however, predicts that ROC curves should look like this:


 

Notice that the main difference is that the ROC curve intercepts above the point (0,0). As you go from the right part of the curve to the left part of the curve its telling you what happens as subjects are getting more and more careful in their responses.  When you get to the Y-intercept (X=0), it represents a situation where subjects are being as careful as they could possibly be, so careful that they accept absolutely 0 lures.  The Y-intercept on the ROC curve thus represents the proportion of "old" items that would be accepted even if subjects were being as careful as they could possibly be.  The only one's you're likely to accept if you're being that careful are ones that you consciously recollect.  And that's Yonelinas's conclusion. A non-zero Y-intercept indicates that there are two memory processes, and the value of the Y-intercept indicates how much recollection is occurring.