Google WikiSkill Gives Agents Persistent Memory Skills

Google Research introduced WikiSkill, a framework that gives AI agents a persistent wiki of failures and successes and turns those notes into reusable Agent Skills, THE DECODER reported on 29 August 2026. On Gemini-3.5-Flash, average performance rose from 49.5 percent to 68.1 percent; Qwen-3.6-27B climbed from 39.4 percent to 63.3 percent. LiveMath for Gemini jumped from 33.0 percent to 72.6 percent, and SpreadSheet from 50.5 percent to 76.6 percent. The system is not true continuous weight learning — the agent writes better instructions for itself after each run.

How does WikiSkill’s three-layer design work?

Inspired by Andrej Karpathy’s LLM Wiki idea, WikiSkill splits the agent workspace into a Raw Layer of immutable execution traces, a Wiki Layer of distilled insights that only grows, and a Skill Layer of procedural instructions that can be rolled back. Each cycle runs an inference agent that produces traces, a Wiki Maintainer that consolidates failure patterns and winning strategies into the wiki, a Skill Proposer that suggests skill edits from the wiki and traces, and a gating step that keeps only updates that improve a held-out validation set. Failed skill proposals still enrich the wiki so later iterations know what not to try again.

That separation is the product point. Skills are executable policies; the wiki is compounding memory. When a skill update hurts scores, the skill rolls back but the knowledge remains. Tang et al. 2026 argue this is why WikiSkill beats prior skill-evolution methods such as Trace2Skill, EvoSkill, and SkillOpt across most model–benchmark pairs in their table.

Which models and benchmarks improved the most?

The study tested Qwen 4B, 9B, and 27B-class models, Gemma-4-31B, and Gemini-3.5-Flash across LiveMath, SealQA, SpreadSheet, OfficeQA, and ALFWorld. Larger models generally gain more from evolved skills, yet smaller models with WikiSkill can match larger models that run without skills. Skills sometimes transfer across model families and can even outperform a model’s self-evolved skills — transfer still needs case-by-case checks. Math and spreadsheet tasks saw the biggest lifts; long-document OfficeQA gains were smaller, and tiny models struggled to execute multi-step search strategies reliably.

Readers tracking agent memory should compare <a href="https://www.promptcrates.com/news/anthropic-claude-shared-memory-chat-cowork">Claude Chat and Cowork’s shared editable memory</a> and <a href="https://www.promptcrates.com/news/meta-evoharness-rl-qwen3-8b-claude-opus">Meta’s EvoHarness-RL results that lifted an 8B model toward Claude Opus</a>. WikiSkill is research infrastructure for skill evolution, not a consumer chat memory toggle, but the product implication is the same: persist what failed, then gate what you ship.

Is this continuous learning of model weights?

No. THE DECODER stresses that the underlying model does not truly learn continuously — still an unsolved problem. WikiSkill is a workaround that compiles experience into text-shaped knowledge and skills. It is more error-prone than real weight updates, the authors imply, but the benchmark deltas show the workaround works. Ablations find that persistent wiki accumulation is critical; without the compounding knowledge layer, skill evolution weakens.

For builders, the practical recipe is clear: store immutable traces, maintain a growing wiki, propose small skill edits, and validate before promote. Roll back skills freely; never wipe the wiki. That design also explains cross-model transfer — skills are instructions, so another model can often follow them even if it did not invent them. Teams wiring agent stacks today can adopt the same pattern without waiting for Google to productize WikiSkill as a named API.

What should product teams file from this paper?

File WikiSkill as a product-update-style research release from Google Research dated around 29 August 2026 coverage. Headline numbers to remember: Gemini-3.5-Flash 49.5→68.1 average, LiveMath 33.0→72.6, SpreadSheet 50.5→76.6, and Qwen-3.6-27B 39.4→63.3. Primary sources: THE DECODER’s explainers and the Tang et al. arXiv paper. Do not market it as online fine-tuning; market it as persistent agent skill evolution with gated rollback. If you already ship tool-using agents, the cheapest experiment is a wiki directory that never deletes failed proposals.

The table in Tang et al. also shows why skill evolution complements model scaling rather than replacing it. Gemma-4-31B with WikiSkill reaches a 54.9 percent average versus 41.3 percent with no skill, while Qwen-3.5-9B climbs from 29.9 percent to 47.4 percent. Those lifts are not uniform across SealQA or ALFWorld, which is why the authors warn against treating any single benchmark jump as a universal agent upgrade. Product managers should pick the task family that matches their workload before citing the headline Gemini average.

Operationally, WikiSkill’s gating step is the difference between a scrapbook of prompts and a release process. Only skill edits that improve a held-out validation set survive; everything else still teaches the wiki what failed. That is closer to continuous delivery for agent policies than to classic MLOps fine-tuning. It also means teams need a stable validation suite — without one, the Skill Proposer has nothing trustworthy to gate against.

Sources

• <a href="https://the-decoder.com/google-gives-ai-agents-their-own-wiki-so-they-can-learn-from-mistakes-and-successes/">Google’s WikiSkill gives AI agents a persistent memory of past mistakes</a> — THE DECODER, 29 August 2026
• <a href="https://arxiv.org/html/2608.27454">WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution</a> — Tang et al., 2026