Feedback Any plans to Multi brain connection

As most of the ML shows its little achievements after some amount of training, these production ready models are way to much hard to integrate to each other.
Say, an agent learns a simple task like walking dynamics during simulated physics environment. And the model is saved. However when this agent is trained to run or jump, it must trained from scratch again.
The previous walking model knowledge becomes useless.

Meantime this kind of training all the time stops what we really need to achieve flexible brain model in ML field.
Is there any ways to solve this bottleneck that integrates all learned small skill together and creates a brand new powerful skill?
I believe these all issues need to be solved inside the dynamic brain system, not the manually changing main model when its need.

Any plan solve these kind of main issue in Machine Learning?