Comment by DenisM

Or better context curation - less lossy compression saving back to context. Maybe even jettisoning part context into an external semantic store instead of conpression. Or placing less data into context to start with.

Or a combination of all those things.