Thanks for the suggestion. I like the idea, but it sounds like it will only work when the objects in question are immutable. The problem here isn’t Character objects (I’m storing char primitives on the nodes) but the actual nodes of the decision tree, which are mutable and thus can’t be reused like that.

I’m going to try to solve the problem by collapsing multiple nodes into one, so that when there are long runs of characters without branches they’ll be stored on a single node. This should hopefully reduce the number of objects by a very large factor (exponentially, in some cases).

I don’t know much about what you’re doing but you might consider using the Flyweight pattern from GoF. They talk about representing characters in a word processor as objects, but using Flyweight to, instead of creating a fresh object for each character, instead creating an object for the first instance of any given character and then just pointing back to the ones you’ve already created. It should work if it’s sufficient for you to point at the character objects from inside your data structure.

If it can be done and you try it, it should improve both your resident size and your performance.