As a devout non-Platonist when it comes to metaphysics, I read the dialogues for those brief dazzling side-arguments which sound for the first time, at least in surviving writings, some problem whose note can then be heard echoing for more than two millenia, into our own time.
I was alerted to one such echo in the characteristically acute and compressed comment of R. M. Hare (Plato, Oxford University: 1982, p. 54): "A lump of flesh does not have desires: my throat does not have desires when I am thirsty; I have them as part of my conscious desires. By the time he wrote the Philebus, Plato was expressing this point very clearly; he proves it from the fact that desire is of something not physically present, which therefore cannot apprehended by the bodily senses, but only envisaged by the mind."
As Socrates sums it up: "the argument, having proved that memory attracts us towards the objects of desire, proves also that the impulses and the desires and the moving principle in every living being have their origin in the soul. . . The argument will not allow that our body either hungers or thirsts or has any similar experience."
Now this point comes back in a very modern form in the distinction between forward chaining and backword chaining in the theoretical foundations of Artificial Intelligence (the distinction has actually much wider application to psychology, cognitive science and the philosophy of science -- but that's another story). If we set aside Plato's denigration of the flesh, and the irrelevancy of phenomenal consciousness introduced by Hare, we see that Plato is giving us a general basis for distinguishing between two functional architectures for action or behaviour. In the one, the organism or system reacts to what is present to the senses; the current state of affairs triggers some behavior -- for example, dryness in the mouth triggers salivation. In general functional terms, this is represented computationally as an instance of a forward-chained "IF - THEN" rule. (Newell and his students at CMU in the '80s built cognitive models based on collections of such rules, and the technology was later applied to expert systems with tools such as CLIPS.) Plato's point is that this cannot represent desire or expectation (or more generally, goal-seeking.) For the desire or expectation is of something which is not true, but nonetheless represented in the mind; that is, the rule which governs or selects a behavior relevant to the desire has the form: "to achieve the goal G, do X" where G is currently false. A system, whether natural or artificial, is "attracted towards the objects of its desire" through a rule which involves a represention of those objects in some form (G) and a way of determining whether in the current state that representation applies.
Rules of this sort are backward-chained: the direction of activation is from the goal to sub-goals: the means (X) by which the goal can be achieved. Whereas in forward-chained rules, activation or causality runs from antecedent to consequent, in backward-chained rules, it runs from consequent back to the antecedent. (Backward chained rules are the basic building blocks of Prolog programs.)
What is so insightful in the Philebus is the sharp differentiation of the structure of intentional from that of non-intentional action, through the crucial role of goal representation.