If you had left off the JVM requirement, I might have suggested ypsilon, since it was built by and for a videogame developer.
But for the JVM, Clojure would probably be the most sensible alternative, if only for the strength of its community right now. You may be able to find lots of people willing to help you out when researching performance problems.
Unfortunately, it may be hard to do an apples-to-apples comparison between Clojure and other JVM implementations of Lisp.
Assuming you are deploying to traditional PCs, rather than phones or small devices, and that your game project is not particularly resource-intensive, your biggest concern probably won't be absolute memory utilization, but hiccups caused by garbage collection pauses. So, in evaluating Lisps, you may want to run some tests to figure out whether there are differences between them in how often garbage collection is triggered, even if it is just caused by different strategies in newing objects. But you'll also have some of the same opportunities to tune garbage collection behavior yourself, by being careful with object creation, or sensible use of weak/soft references, or even just using command line tweaks JVM gc options when you start up the app.
You could also try building some benchmarks that are useful to you, and test whether different JVM Lisps exhibit different characteristics. A number of people have invested research into profiling Lisp; see, for example, http://openmap.bbn.com/~kanderso/performance/ and consider whether some of the approaches linked there would be relevant for your application.
In at least one set of tests, http://per.bothner.com/blog/2010/Kawa-in-shootout/ , Kawa performed somewhat better than Clojure on many tests. If that's a major consideration, you may want to at least explore whether similar benchmarks apply to your case or not.
It doesn't sound like you are overly concerned with this, but if a (nearly) complete common Lisp implementation were important to you, Armed Bear Common Lisp might be a better place to start than either of these. I like it just for the name, but that's probably not a good enough reason.
The biggest challenge of answering a question like "which is best for..." is the implicit, unstated "by some measure of best." For me, the strength of a community around a technology is often at least as important as the performance considerations, since I rarely develop in a complete vacuum.