We present a longitudinal study measuring whether consensus-validated institutional memory produces measurable, cumulative improvement in AI agent performance across sequential task executions. Using (S)AGE, a 4-node BFT-backed institutional memory layer, and CipherForge Labs, an 11-agent organization with 3-line prompts and zero domain expertise in their instructions, we conduct three experiments: (1) a difficulty sweep across 9 tiers establishing baseline performance, (2) 10 sequential runs at maximum difficulty with a red team feedback loop measuring longitudinal learning, and (3) a 20-run control arm with expert-crafted prompts and no institutional memory. Red team assessed difficulty increases from 0.8 to 3.0 over 10 sequential runs (Spearman ρ = 0.716, p = 0.020), with zero prompt changes and two knowledge-routing fixes applied during Phase 2. The control arm shows no learning trend (Spearman ρ = 0.040, p = 0.901). Cross-sectional means are statistically indistinguishable between arms (Cohen’s d=−0.07), confirming that the contribution is the learning dynamic, not absolute performance level. This paper builds on our prior finding (Paper 3) that 3-line prompts with institutional memory produce a fully autonomous security research loop. Here we ask: does the loop get better over time?
Dhillon Andrew Kannabhiran (Fri,) studied this question.