Didn’t openclaw switch to vector based because it used way less tokens as it always loaded all memories? Seems way more efficient