It does, exactly - it might still be a viable option depending on how you compress it.
Every second token IMHO is still better than removing every second word or sentence.
Also summarizing will remove some stuff that from users point of view might be important but the model doesn’t know what is and what is not - so it most likely takes out a statistically most likely - but who is average…
Newer methods like self aware attention work better in that regard - since the model can learn something about the users preferences over time.
Combining that ability with previously extracted memory snippets (Memory Augmented Generation / MAG) like ChatGPT does already is pretty smart.
We’ll get there (and beyond).