I am not sure exactly what you mean, but there are several ways in which to save space:
1. Whenever a column receives a very low weight (near zero), you can move that weight to the column with, for example, the least objective value and throw the zero-weight column away.
2. You can aggregate columns, too: for example, you can define one new column by taking any convex hull of some of the columns.
Whether you - in the end - have saved any substantial amount of space depends on the data in the problem.
Luis: The proposed ideas of Michael are very good suggestions. In the case of 1., you can also consider to move a column according to the number of times that this column was part of the optimal basis in the solution of the diferent reduced master problem solved during the iterations of the method. With my best regards, Víctor A.
Louis, do you really mean "column generation" instead of "aggregation"?
Because, Dantzig-Wolfe that you mention is basically a decomposition technique known as "column generation", not "aggregation of columns". And yes, it was devised so as to save memory for large LP problems, however to be effective, your problem must have a rather special structure (e.g. if your constraint matrix is a block-angular matrix)
Another solver (+ Python-based mathematical modelling language) for column generation (that constructs the Dantzig-Wolfe restricted master problem for you) is DIP (with Dippy)