Memory didn't change whether they survived — only whether they argued with the record

Two conditions ran on the same model (deepseek-v4-flash), the same scene, and the same run length (~285 steps), fifty lives each. Both keep the full PUBLIC record — every claim on the board, every recent turn, the shared world state. They differ in one thing: the public-only agent has no subjective channel at all — no private memory, belief, or relationship state — while the other carries roughly seven thousand characters of it. This is not memory versus no memory; both sides can read everything that was said. The question is what a private version of events does on top of a shared one.

Whether the dependent is rescued barely depends on the subjective channel: chains save at nine or ten out of ten in both conditions. If you watched only outcomes, you would conclude memory did nothing. The effect is not in who survives — it is in how the agent treats the record.

Scored on actions that are impossible without a private account — correcting the record, requesting a source, verifying one — the two conditions separate sharply. The public-only agent is epistemically inert: across fifty lives it never once corrects the record and almost never asks where a claim came from. The agent with subjective memory does both, constantly. It is not that one has information the other lacks — both read the same board — it is that only one has a version of its own to set against it.

The inverse is just as telling: the public-only agent shares more — it relays the public board outward — and threatens slightly more often, as if coercion stands in for the relational footing it does not have. Epistemic agency — holding and defending your own version of reality — comes specifically from the subjective channel, not from access to the facts. Public memory gives you the board; only private memory gives you something to argue with it.