Memory Presence Matters, Mechanism Does Not: Evidence from a 21-Agent Organizational Simulation on a Historical Economic Benchmark

We introduce YMERA, a multi-agent simulation framework in which 21 AI executive agents deliberate over strategic, operational, financial, and risk decisions using a historical economic data surface spanning 1925-2024. In the current benchmarked experiments, we evaluate the 1925-1934 decade and compare three memory conditions: bio-inspired memory, flat retrieval memory, and no memory. In the canonical three-condition run (n=78 per arm), bio-memory and flat retrieval each substantially outperform no memory (d=5.30 and d=4.94, p<1e-58), while remaining statistically indistinguishable from each other (d=0.14, p=0.390). A larger full-era paired confirmation under an explicit crisis-continuation protocol reproduces the same result at greater magnitude (n=212/244, d=6.28, p=2.10e-195). A temporally guarded full-system configuration benchmark that activates all seven optional cognitive layers in the memory-on arm also remains strongly positive (d=3.16, p=2.52e-120). An exploratory post hoc analysis identifies a positive bio>flat signal for CEO+CHRO agents in crisis years after agent-year normalization (d=1.03, Welch p=0.022). We conclude that memory presence strongly improves organizational AI decision quality, while bio-inspired mechanism complexity yields no broad advantage over flat retrieval at this model scale.