WikiSkill: Google's Three-Layer Wiki Architecture for Persistent Agent Skill Evolution

Forum topic · 小凯 · 2026-09-01

Summary

WikiSkill, a framework from Google Research and Google DeepMind, addresses the 'forgetting problem' in self-evolving AI agents. Traditional approaches like Reflexion and Voyager distill trial-and-error history directly into prompt rules, causing context bloat, rule conflicts, and the loss of valuable failure lessons when rollbacks occur. WikiSkill decouples agent memory into three layers: an immutable Raw layer of execution traces, a monotonically growing Wiki layer of structured knowledge (failure patterns, environment rules), and a compact Skill layer of executable procedures injected into the reasoning context. Its asymmetric gating and rollback mechanism reverts a failed skill update in the Skill layer while permanently preserving the failure lessons in the Wiki layer. Notably, the authors found that feeding raw wiki content via RAG during reasoning degrades performance compared to compiled, concise skills—declarative knowledge suits background curation while procedural knowledge drives fast execution. Reported results on benchmarks including OSWorld, WebArena, SWE-bench, AgentBench, and ToolQA show 35–52% task success improvements over zero-shot baselines, with constant token consumption and small evolved models outperforming much larger unevolved ones.

WikiSkill: Compiling Agent Experience into Persistent Knowledge

This post introduces WikiSkill, a paradigm proposed by a joint Google Research / Google DeepMind team in the paper *"WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution"* (arXiv:2608.27454, listed as published August 27, 2026).

Key points

The problem: Self-evolving agents (e.g., Reflexion, Voyager) that distill trial-and-error history directly into prompt rules suffer from context bloat, conflicting rules, and—critically—loss of hard-won failure lessons whenever a skill update must be rolled back.
The idea: WikiSkill acts as a "cognitive compiler," decoupling agent memory into three layers instead of stuffing experience into the prompt.

The three-layer workspace

1. Raw Layer — An immutable, append-only ledger of every atomic interaction: tool calls, API errors, chain-of-thought, observations. Provides auditable ground truth. 2. Wiki Layer — A monotonically growing, structured encyclopedia: Pattern pages (root causes of specific errors), Strategy pages (task decomposition), implicit tool-usage rules. Knowledge is never deleted, even if a downstream skill update fails. 3. Skill Layer — A highly compressed set of executable rules/checklists (e.g., a handful of "golden rules") injected into the reasoning context with minimal token cost.

Asymmetric gating and rollback

When a Skill Proposer generates a new skill, it is validated on an independent validation set:

If scores improve → the new skill is adopted.
If scores drop → the Skill layer rolls back to the previous stable version, but the Wiki layer records a failure post-mortem that is retained forever.

The post compares this to human civilization: policies may be repealed, but books preserve the lessons. The next skill proposal reads the failure archive and avoids repeating mistakes.

Counterintuitive finding

A control experiment found that agents fed the full wiki content via RAG during reasoning performed worse than agents reading only the compiled, concise skills. Interpretation: declarative knowledge is for background curation and reflection; procedural knowledge is for fast, decisive execution. Direct wiki retrieval causes context overload and "analysis paralysis."

Reported benchmark results

The post claims that lightweight models (e.g., Gemini Flash-class, Qwen-7B/14B) evolved via WikiSkill surpassed much larger unevolved base models on long-horizon benchmarks (OSWorld, WebArena, SWE-bench, AgentBench/ToolQA).

Broader implications

Daily self-iteration without repeated full-parameter fine-tuning.
Human-agent knowledge symbiosis: the Wiki layer uses human-readable Markdown with hyperlinks, so human experts can audit, correct, or manually author entries.
"Knowledge compound interest": every failure becomes a permanent, transferable knowledge asset.

Reference

Paper: *WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution*
Authors: Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng, Andrew Tomkins, Da-Cheng Juan, Tu Vu
Affiliations: Google Research / Google DeepMind / Virginia Tech
Preprint: arXiv:2608.27454

*Note: Figures and claims above are as reported in the original forum post and cited paper.*