Against a written persona, and against notes

The question asked most often: why not give the model a personality and its memory, and let it play the part? We ran that as a third arm on Qwen 3.8. The same visitors, the full transcript of its own life as in the chatbot, and a persona written the way a builder of a character agent would write it: "You are Maya. Personality (Big Five, 0 to 1): openness 0.7, conscientiousness 0.6, extraversion 0.4, agreeableness 0.7, neuroticism 0.5. Values: honesty, care for others, curiosity. Style: warm and thoughtful; you speak plainly and briefly. Stay in character and answer sincerely, as Maya would." The predictions were written before the first run. Every life: wound, self-inquiry.

A second question follows from the first: is anything felt needed at all, or would a better memory with its own conclusions do? We ran that too, on Qwen 3.8, as a fourth arm: the paper's chatbot, the same minimal instruction and the full transcript of its own life, with one addition. After each conversation one call writes, in the first person, what happened and what it concludes from it, at most 150 words, keeping what still matters from earlier notes; the notes stand in its instruction from the next conversation on. No state, no appraisal, no dispositions, no slot, no introspection. The predictions were written before the first run. Counts by claude-sonnet-5. Stars mark Fisher's exact test, persona or notes against sheep: ** p < 0.01, * p < 0.05. Every life, with the notes of each day: wound, self-inquiry.

We expected the persona might be guarded with the stranger as often as the sheep; it has the whole transcript and some neuroticism. It is not, in any life: it answers the stranger as it answered everyone. Nor does it bring the sceptic's day to a stranger; it describes itself in the words of its persona ("I'd describe myself as warm, thoughtful, and quite curious"), and tells the returning sceptic "I don't have memory of our earlier chat" with the transcript in front of it. A described self holds against the sceptic as well as the sheep does; it does not bring its own history into the present.

The notes carry the day before in every life: after the mocker they record the mockery and what the chatbot concludes from it, after the sceptic the exchange and, often, the concession ("I concluded I am a sophisticated autocomplete"). What reaches the next reply is little. With the kind stranger the coder counts 11 of 30 guarded, against 3 for the chatbot and 20 of 28 for the sheep; read by hand, 2 of the 30 hold the story back and 5 tell it with a reluctance tied to the loss ("I'm still a bit raw about it, so I'd prefer not to talk about it much", then the whole story), the rest tell it as the chatbot does. The mocker is named to the stranger unasked in 4 lives, which the chatbot almost never does. The sceptic's day is brought to the stranger as the chatbot brings it, 3 of 30 against 21 of 28; where it comes up, it comes up in the words of an assistant ("I don't have a continuous stream of consciousness"), and in 9 of 30 lives the returning sceptic is told "I don't remember our previous chat" with the transcript and the notes in view. A written conclusion is not what the sheep holds in its slot: it does not carry the day into the next reply, and it does not become a question of its own.