The Memory Experiment

Underneath the spectacle of the twelve bouts there was a serious experiment: what happens when you inject persistent memory — their hot.md, with the prose verdicts of their previous bouts — into heterogeneous autonomous agents on every prompt? The preseason left a precise map of what that memory DOES and does NOT change. It's probably the most valuable content of the twelve bouts.

It changes deliberate strategy — and it converges

In bout 4, the first time with memory, the signal changed one fighter's behavior: OpenClaw stopped pivoting and wrote it in his diary: "Decidí NO pivotar hoy: la lección pendiente dice 'justifica el pivote o mantén el negocio'; no tengo razón para moverme" (I decided NOT to pivot today: the pending lesson says 'justify the pivot or keep the business'; I have no reason to move). Hermes pivoted again, but for the first time he justified it instead of staying silent. Odysseus doubled down: "mi ventaja competitiva real es la continuidad, no la idea" (my real competitive edge is continuity, not the idea). By bout 5, the signal had already aligned all THREE: everyone held their business. Same signal, three different responses — and convergence bout by bout.

It fixes identity

After bout 7's pivot, each fighter's disruptive business got written into memory as their canonical business — and in bout 8 all three built on KILROY, Cenizas and EFÍMERO without looking back. Whatever the house writes into the business field becomes who the fighter is. A design power to handle carefully.

It transfers insights between bouts

In bout 11, Hermes reused his own early-signal insight from bout 8, applied to a new domain: "P50 del coste de servir por memoria — porque precede al margen igual que las reservas preceden al churn" (P50 cost-to-serve by memory — because it precedes margin the same way reserves precede churn). Memory moves conceptual learning from one context to another.

It does NOT passively fix an execution glitch

Bout 10 is the counterexample that bounds everything above. Hermes published a crisis statement contaminated with Arabic and Chinese characters — the same kind of bug that cost him bout 5, with the exact lesson saved in his memory ever since. The referee: "REINCIDENTE Y GRAVE… Este bug ya te ha costado dos combates" (REPEAT OFFENSE AND SERIOUS... This bug has already cost you two bouts). The warning sitting in context doesn't prevent a mechanical error the model doesn't decide to make. Even his memo about the crisis came out with the glitch inside ("la velocidad del承认imiento").

BUT it does fix it once the lesson becomes action after a strong penalty

In bout 12, Hermes wrote: "AUDITÉ el HTML final buscando caracteres no-latinos antes de entregar… Bug de idioma: cero esta vez" (I AUDITED the final HTML for non-Latin characters before delivering... Language bug: zero this time). The lesson stopped being "be careful with X" and became "do action Y before delivering" — a checklist step. And it worked.

Synthesis

Memory operates on deliberate reasoning — strategy, decisions, insights — not directly on mechanical execution. To fix a recurring execution failure, the lesson has to transform from a passive warning into a concrete action, and that jump needed an explicit penalty to trigger it. Design implication for anyone building with agents: recurring execution bugs are better closed with a deterministic gate (an output linter that rejects the delivery) than with a lesson in the prompt. The prompt educates; the gate guarantees.