As many have pointed out: Gromov introduced it, and we continue to use it today because it's an incredibly useful tool. I've never spoken to Gromov about why he introduced it (who knows how great mathematicians come up with great ideas) but I can try to give some (probably historically false) motivations as to why someone might have come up with the notion. For instance, if Gromov hadn't discovered it, you might have come up with it as follows:

(1) First, complex geometry--if you like, you can think of algebraic geometry--has a lot of rigidity. The fact that we can even give a discrete count to sub-objects (like how many curves pass through n fixed points) is special -- the question takes on a totally different nature in more flimsy geometries.

Now, is there a way to relax the background of complex geometry, and still come up with a useful, fun theory? For instance, how necessary is the integrability condition on J (the complex structure) to still make sense of curve-counting?

What Gromov showed is that if the complex structure is `tame' in the sense that one has a compatible symplectic form, questions about curve-counting can still have answers. Really, the difference between a pseudoholomorphic curve and a holomorphic curve isn't in their definitions, it's in the nature of J in the target. Relaxing the J from "integrable complex structure" to "complex structure tamed by a symplectic form" is the generalization that's happening.

(1') Put another way, we already had a famous 2-out-of-3 principle recognizing the relationship between Riemannian, complex, and symplectic structures on a vector space. Studying curves on complex projective varieties take on rigidity, in some sense, because we study maps between manifolds with Kahler structure (manifolds both symplectic and complex, and further, each structure is integrable--in that the Nijenhaus tensor vanishes, and omega is closed.) It's natural to ask whether we can still find interesting structure in the 2-out-of-3 world by studying manifolds whose tangential structures are compatibly Riemannian, complex, and symplectic, but which do not satisfy a global condition like integrability of J or closed-ness of omega. And when you get rid of the integrability of J, it turns out that you can find such a structure on any symplectic manifold. (In fact, once you fix $\omega$, there's a contractible space of compatible $J$. That's why pseudoholomorphic curves can be applied widely in the symplectic world.

(2) There might be another motivation from physics. In mirror symmetry, one predicts the existence of mirror Calabi-Yau manifolds. A field theory that relies on the symplectic structure of one manifold should correspond to a field theory that relies on the complex structure of the mirror. And the correlation functions count J-holomorphic curves in the symplectic manifold. Historically though, I'm not sure if physics alone would be able to motivate the study of these field theories on just symplectic manifolds with almost-complex structure, as opposed to Calabi-Yaus. Somebody with more background could probably comment on this.

As many have pointed out: Gromov introduced it* wrote a seminal paper utilizing it, and we continue to use it today because it's an incredibly useful tool. I've never spoken to Gromov about why he introduced it (who knows how great mathematicians come up with great ideas) but I can try to give some (probably historically false) motivations as to why someone might have come up with the notion. For instance, if Gromov hadn't discovered it, you might have come up with it as follows:

(1) First, complex geometry--if you like, you can think of algebraic geometry--has a lot of rigidity. The fact that we can even give a discrete count to sub-objects (like how many curves pass through n fixed points) is special -- the question takes on a totally different nature in more flimsy geometries.

Now, is there a way to relax the background of complex geometry, and still come up with a useful, fun theory? For instance, how necessary is the integrability condition on J (the complex structure) to still make sense of curve-counting?

What Gromov showed is that if the complex structure is `tame' in the sense that one has a compatible symplectic form, questions about curve-counting can still have nice answers. Really, the difference between a pseudoholomorphic curve and a holomorphic curve isn't in their definitions, it's in the nature of J in the target. Relaxing the J from "integrable complex structure" to "complex structure tamed by a symplectic form" is the generalization that's happening.

(1') Put another way, we already had a famous 2-out-of-3 principle recognizing the relationship between Riemannian, complex, and symplectic structures on a vector space. Studying curves on complex projective varieties take on rigidity, in some sense, because we study maps between manifolds with Kahler structure: manifolds both symplectic and complex, and further, each structure is integrable--in that the Nijenhaus tensor vanishes, and omega is closed. It's natural to ask whether we can still find interesting structure in the 2-out-of-3 world by studying manifolds whose tangential structures are compatibly Riemannian, complex, and symplectic, but which do not satisfy a global condition like integrability of J or closedness of $\omega$. And when you get rid of the integrability of J, it turns out that you can find such a structure on any symplectic manifold. (In fact, once you fix $\omega$, there's a contractible space of compatible $J$. That's why pseudoholomorphic curves can be applied widely in the symplectic world.)

(2) There might be another motivation from physics. In mirror symmetry, one predicts the existence of mirror Calabi-Yau manifolds. A field theory that relies on the symplectic structure of one manifold should correspond to a field theory that relies on the complex structure of the mirror. And the correlation functions count J-holomorphic curves in the symplectic manifold. Historically though, I'm not sure if physics alone would be able to motivate the study of these field theories on just symplectic manifolds with almost-complex structure, as opposed to Calabi-Yaus. Somebody with more background could probably comment on this.

*As I learned from Antoine and Dmitri, there were previous works utilizing pseudo-holomorphic curves. For instance:

A.Nijenhuis, W.Wolf. Some integration problems in almost-complex and complex manifolds, Ann. Math. 77 (1963),

J. Eells and S. Salamon. Twistorial construction of harmonic maps of surfaces into four-manifolds. (1985).