Episode compression at task boundaries + pressure compression as fallback for non-episode context gives best of both approaches.
https://randomlabs.ai/blog/slate