--- title: "Can AI Agents Learn From Experience Now? Wikis, Red Teams, and Failure Mining — the Papers Read Inside China" date: 2026-08-29 originalDate: 2026-08-29 originalTitle: "AI学会复盘了，经验还值几年钱？" issue: "Batch 1" description: "Agent experience compiled into evolving wikis, red-team agents that learn from attacks, small-model failures tutoring big models — plus Google's ReasoningBank." tags: - "agents" - "memory" - "red-teaming" - "experience reuse" - "weak-to-strong" papers: - id: "2608.27454" title: "WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution" url: "https://arxiv.org/abs/2608.27454" - id: "2608.27439" title: "RedEvoAgent: Automatic Red-Teaming Agent with Experience-Driven Skill Evolution" url: "https://arxiv.org/abs/2608.27439" - id: "2608.27455" title: "CritICL: Inference-Time Weak-to-Strong Generalization from Small Language Model Failure Modes" url: "https://arxiv.org/abs/2608.27455" sources: - label: "Google Research blog (ReasoningBank — agents learning from experience; project page linked from the blog index)" url: "https://research.google/blog/" faq: - q: "What do these papers add up to?" a: "A shift from per-task learning to compounding experience: WikiSkill separates raw execution traces from a persistent, evolving wiki of distilled skills; RedEvoAgent gives attack-side agents the same compounding; CritICL shows small models' failure modes have structured, scale-transferable regularities usable to steer big models at inference time — no re-generation, no external verifier. Google's ReasoningBank (blog, not arXiv) drives the same direction: reasoning experience stored as retrievable memory." - q: "Does this mean human experience is worthless?" a: "The Chinese take's three-layer answer: information (already dead), skills and process (under attack now — this batch), judgment and accountability (still safe, not because AI can't but because 'who answers for it' is not a machine question). Its updated maxim: stock trust, responsibility, and the willingness to sign, not experience." - q: "What are the known failure modes?" a: "All results are lab-environment. The take's own caveats: real-world noise can compound wrong experience (an agent accumulating in the wrong direction compounds confidently); and human-derived experience carries copyright, privacy and compliance knots that synthetic wikis sidestep — until they don't." --- Why can't ChatGPT remember what you taught it last week? For years that was the wage-earner's consolation about AI. A late-August arXiv batch plus a Google Research project aim to retire the consolation: agents that keep notes. Our Chinese-language column's roundup ran under "AI has learned to do post-mortems — how many years is experience still worth?" This digest translates the reading; paper claims follow their abstracts. ## The claims - **WikiSkill (2608.27454):** a two-tier architecture — raw execution experience stays raw; distilled knowledge goes into a persistent wiki that co-evolves with a skill library, accumulating across iterations. This round's pitfalls are next round's detours. - **RedEvoAgent (2608.27439):** the same compounding, weaponized — a red-team agent that distills attack experience into skills for jailbreaking product-grade agent environments, where triggering an agent is no longer "says a rude thing" but "executes a dangerous action with real tools." - **CritICL (2608.27455):** small models within one family exhibit structured failure modes that transfer across scale; those failure catalogs can steer large models at inference time — improving the strong with the weak's mistakes, without repeated generation or external verifiers. - **Google Research's ReasoningBank** (blog): agents storing reasoning experience as retrievable memory — big-lab and paper-lab traffic converging on the same intersection. ## How the Chinese column reads it The anchoring image is corporate, not sci-fi: **the project post-mortem document** — except the humans' version sits unread while the agent's version is genuinely consulted ("it took companies a decade to build the review habit; AI installed it with one paper"). On the red-team paper, the chill the column isolates: experience doesn't check motives — the defense's AI writes its wiki, the attack's AI writes its wiki, two thickening notebooks shelling each other with our data in between. CritICL gets the inversion prize: failure used to be training's garbage; now it is mined as ore. And the column's three-layer depreciation schedule for human moats: information gone, skill-and-process depreciating quarterly, judgment-and-accountability intact — for reasons of responsibility, not capability. ## Cold water Lab environments, all of them. Compounding cuts both ways — an agent accumulating confidently in the wrong direction is a bigger liability than a goldfish. And the column names the clock: the gap between these papers and your workstation is measured in quarters, which is the only window the advice-buying public gets. ## What to watch Memory becoming a product surface (whose wiki? which sandbox? portable across vendors?); red-team compounding forcing a re-price of agent permissions in production; and whether "experience" claims start appearing in model marketing — the moment that word reaches a landing page, this batch becomes a spec sheet. ## Sources - [WikiSkill (arXiv:2608.27454)](https://arxiv.org/abs/2608.27454) - [RedEvoAgent (arXiv:2608.27439)](https://arxiv.org/abs/2608.27439) - [CritICL (arXiv:2608.27455)](https://arxiv.org/abs/2608.27455) - [Google Research blog (ReasoningBank)](https://research.google/blog/) > **Provenance & disclosure.** Originally published in Chinese on our WeChat channel on 2026-08-29 ("AI学会复盘了，经验还值几年钱？"); drafted with AI assistance under human editorial direction. Translated to English on 2026-08-29 (AI-assisted, human-reviewed). Paper claims follow the arXiv abstracts as listed by the column; ReasoningBank is cited from Google Research's blog, not a paper, and is labeled as such. This is translated commentary — not a SigPulse measurement. Our first-party measurements live in the [dispatches](/posts/) and the [/data/ ledger](/data/).