Self Without a Transcript: Reading, changing and reusing a conditional policy

What must a model remember to edit a policy after its original words are gone? A four-row continuous action memory passed every registered single-edit test under familiar wording, with 3,072/3,072 correct Gemma choices. Reusing the edited table passed only three of nine two-edit sequences as the branch roles moved between rows. Two histories can even have identical current actions and require opposite choices next. Given the trigger mask, a nine-parameter controller learned from externally scored Gemma choices to carry the role of the first row in one bit. It resolved all 9,216 paired tests in each of three seeds, matching explicit role supervision. This evaluation reuses 16 policies and 196 distinct native inputs. The registered advantage over a no-bit control failed at the fourth edit. In an exhaustive follow-up, all four answers were correct on 100% of trajectories with either learned bit and 12.5% without it. With three roles, six assignments shared one table. Two of three reward-trained controllers resolved all six through a three-bit channel. The third resolved five. Rewriting equivalent rules in raw Gemma slots remains open. Two learned updates gave 252/512 correct choices, and the independently encoded target rule and zero input each gave 256/512. Correspondence: [email protected]