ORALMEM: Post-Witness Collective Memory in Large Language Model Societies

Societies of large language model (LLM) agents are often given persistent histories, retrieval systems, summaries, or shared stores. These tools make agents easier to study, but they hide a basic social question. Can a population keep an event alive through conversation after the original witnesses no longer remember it? We introduce ORALMEM and the Witness Extinction Protocol (WEP), which removes target-derived memories from every original witness and then probes only non-witnesses. In 700 controlled mechanism runs, recurrent oral retransmission sustains post-witness memory under bounded FIFO memory. At the baseline setting, recurrent relay reaches PWR@50 of 0.397, compared with 0.105 for one-hop spread. The gap is positive in all ten paired worlds. Live GPT experiments show why this persistence is social and policy-dependent. A forced four-hop chain preserves a complete event after the witness forgets, while an unforced chain stops when an informed carrier chooses silence. Repeated one-root reports rarely create the illusion of independent sources. The more interesting pattern is a split between evidence accounting and use. Models can report one root source and still act on a repeated report, with action rates that depend strongly on framing. Together, the results present post-witness memory as a pipeline in which availability, retelling policy, semantic retention, provenance, confidence, and action can diverge.