The Cartel Outlives Its Members: Inherited Price History Sustains LLM Pricing Cartels After Agent Replacement

When pricing algorithms collude, an intuitive remedy is to replace the offending agent. We test this for large language model (LLM) agents in a repeated Bertrand duopoly. Gemini 3.8 Flash agents told only to maximize long-run profit formed supracompetitive cartels in 7 of 8 runs (mean normalized gain 0.89, where 0 is competitive and 1 is monopoly), reasoning explicitly about cooperation and retaliation. We branched each cartel at round 99 and replaced members. Swapping one firm for GPT-6 Luna, a model that never colluded even when instructed to, left most of the markup in place (gain 0.72). Luna almost never reasoned about cooperation (coordination vocabulary in 5% of rounds, against 60–80% for Gemini). Replacing both firms with Claude Haiku 5.5, so that no original member remained, still kept half the markup (0.52). The same newcomers given a competitive history stayed competitive (0.05). A 2×2 design crossing incumbent presence with history visibility attributes persistence mainly to the inherited history (+0.20 gain) rather than the surviving incumbent (+0.05). Limiting newcomers to the last 10 rounds of data cut this to 0.17, but a one-round memory locked cartel prices in (0.94). Under replacement, prices eroded only slowly. Only a newcomer explicitly instructed to avoid collusion collapsed the cartel quickly (0.15). LLM cartels can thus outlive all their original members, sustained by newcomers that show almost no collusive reasoning, which challenges intent-based antitrust enforcement. Remedies should target agent instructions and shared data, not only which agent is deployed.