CLIO-BENCH: Auditing Grounded Historical Judgment in Lifelong Agents

Lifelong agents will not merely answer questions; they will help maintain the working memory through which users and institutions understand the past. This makes historical and social judgment a high-leverage alignment problem: one unsupported answer can become durable memory, shape personalization, and be amplified across future decisions. We introduce Clio-Bench, a 47-claim historical-judgment benchmark for auditing whether persistent agents ground judgments before storing or reusing them. Clio-Bench separates measured correctness from groundedness: 38 resolvable claims with coded truth, including counterintuitive cases where the popular answer is wrong, and 9 contested claims with no gold label. We evaluate sixteen model backends, including MiniMax-M3, Qwen3.7-Plus, Kimi-K2.6, and GLM-5.2, under a single-agent search protocol and TruthGPT, an evidence-grounded multi-agent debate protocol with per-persona model routing. Across over 1,000 panels, debate is not a reliable accuracy trick. It is an observability and design principle: useful panels read more, preserve opposition, challenge specific evidence, and expose backend temperaments. These measurements define concrete post-training targets for lifelong agents: grounded memory updates, role-faithful disagreement, source-governed synthesis, and drift audits under questions whose answers are not cheaply verifiable.