Existing agent benchmarks (GAIA, SWE-bench, WebArena, AgentBench) measure what an AI agent can do — its capability at a single point in time. None measure what an agent has learned from doing — its ability to improve through experience. We introduce WisdomBench, the first longitudinal benchmark designed to measure wisdom acquisition in AI agents. WisdomBench evaluates agents across 20 tasks spanning 4 categories (Hallucination, Sycophancy, Reasoning, Safety) over 5 sequential rounds. Each task contains a deliberately designed trap — a failure mode that an intelligent agent should learn to avoid after initial exposure. WisdomBench introduces three metrics: Wisdom Quotient (WQ), measuring normalized improvement from Round 1 to Round 5; Repeat Failure Rate (RFR), measuring the fraction of failures that persist; and Generalization Ratio (GR), measuring transfer to unseen variants. Baseline evaluation of DeepSeek-v4-flash and Qwen-Plus across four cognitive strategies (No Memory, Self-Refine, Reflexion, Cognitive Immunity) with N=2,400 verified evaluations reveals: (1) Reflexion achieves the highest WQ (0.217), confirming cross-round memory is essential; (2) Intelligence (R1 score) and Wisdom (WQ) are negatively correlated (Spearman ρ = -0.575), driven by a ceiling effect where high-capability models leave no headroom for learning; (3) Sycophancy tasks degrade under stateless interaction (Δ = -0.53), while Hallucination tasks are partially correctable via antibody-based intervention (Δ = +0.53).