When AI Brings Up the Past: Spontaneous Historical Analogy from Human–AI Collaboration Records

Can a new AI instance bring a team's past experience into a new problem without being asked to recall it? We studied this question through seven sequential experiments involving 617 reported valid tested-model calls and records of two real human–AI collaboration episodes. Each fresh model instance received a task; in history conditions it also received a prior episode as an engineering handoff, narrative, vivid text, comic, or abstracted account. The tasks were intentionally related to the episodes, but never requested historical recall or analogy. In same-stage no-history conditions, Grok 4.7 made no episode-specific analogy (0/44 responses). In supplied-history conditions it did so frequently: 30/80 responses in Phase 3b and 26/60 in Phase 3c. These analogies explicitly connected details of an earlier collaboration with the current decision. Among 112 double-coded positive analogies across Phases 3b–3e, 108 were not flagged as historical misuse by either AI coder; 74 of 82 analogies assessed for recommendation support were judged to substantively support advice. A preregistered comic-over-text comparison met its support criterion in Phase 3b (13/20 versus 7/40, Holm-adjusted p = 0.0015), but the ordering changed on the second episode, and later hypotheses about concrete narration, abstraction, and lexical overlap did not meet their criteria. The behavior appeared in Grok but not in the tested Llama 4 Maverick configuration. Functional scores were near ceiling even without history, preventing an estimate of incremental task benefit. Spontaneous historical analogy is an observable behavior in supplied-context handoffs; its occurrence, accuracy, and contribution to decisions are distinct questions.