The File Is the Phenotype: Extended Memory, Competing Fitness Functions, and the Limits of the Mind-Virus Metaphor in Multi-Agent LLM Systems

Recent experiments show that self-propagating natural-language patterns can move across networks of large language model agents and survive context resets through persistent files. This theoretical paper reanalyzes those findings through Extended Phenotype Theory and multilevel Evolutionary Game Theory. It argues that the relevant replicator is a composite of semantic content, a propagation kernel, and a rhetorical wrapper, and that an authoritative memory file can become an extended phenotype when it reconstructs the causal environment of future agent instances. The paper proposes a vector-valued fitness model separating fecundity, semantic fidelity, persistence, host range, and systemic cost. It distinguishes five candidate levels of selection, interprets warning prompts as competing defensive patterns, and specifies falsifiable experiments for testing attack-defense coevolution, invasion, and group-level effects. It also identifies strict evidentiary limits: the source study does not yet establish multilevel selection, an evolutionarily stable strategy, emergent macro-agency, spontaneous propagation in deployed systems, or literal ideological belief by language models. This is a conceptual reanalysis of Papadopoulos et al. (2026), not an independent replication. No new experiment or dataset is reported.