Question: Gradient projection cont'd

Dear all,
Some time ago I asked a question on maple primes concerning handling the numeric solutions of dsolve procedure.

They sent me to the discussion on option remember and indeed this helped a lot.
However there is another issue.

The method I am trying to implement requires iterative calls to dsolve/numeric with one of the functions defined from the previous solution. As a result after 3-4 steps I get the very long function:
Say ODE is diff(x,t)=f(x(t),u);
and I get solution for some u[0]=const.
Then I calculate u[1](t)=u[0]+ff(x(t),u[0])
and use it to solve the same ODE again:
diff(x,t)=f(x(t),u[1](x(t),t)) and now this ODE is a function of the previous solution through u[1]. With more iterations the problem becomes even more complicated.
Is there any workaround for this issue?
As far as I think, the best solution would be to have Maple to define u[1] as function of t only, since it 'knows' previous solution. However I do not have any idea how it can be made, since the solution for dsolve is in the form of procedure from dsolve/numeric...
I would be very grateful for any help or advice, as this is necessary for the progress of my research.