important caveat: this isn't universal GPT-5.2 did show quality drops when assistant history was removed entirely. stronger models apparently extract more signal from their own prior context so the right approach isn't "always strip everything." it's selective filtering. which

→ View original post on X — @godofprompt