What are the three layers in WikiSkill?

The Raw Layer stores full execution traces, the Wiki Layer holds distilled insights that never reset, and the Skill Layer contains active instructions that can be rolled back if they hurt performance.

Does WikiSkill work for small models?

Smaller models like Qwen-3.5-4B struggled with long multi-step tasks, but they still improved. Small models using WikiSkill could match larger models without it.

Large Language ModelsAI Safety & AlignmentTHE DECODER1h agoAug 29, 2026, 22:03 JST2 min read

Google's WikiSkill gives AI agents a memory of past mistakes

Send on LINE

Key takeaway

Google's WikiSkill gives AI agents a persistent memory of past mistakes.
It outperformed other methods across five benchmarks, boosting average accuracy by up to 18.6 percentage points.
This workaround helps agents learn without continuous retraining.

3 Key Points

What happened
Google researchers introduced WikiSkill, a framework that gives AI agents a persistent memory of past mistakes and successful strategies. It organizes an agent's workspace into three layers: a Raw Layer for execution traces, a Wiki Layer for distilled knowledge that never resets, and a Skill Layer for procedural instructions that can be rolled back if a change hurts performance.
Why it matters
WikiSkill consistently outperformed other skill evolution methods in the study. On average, it boosted Gemini-3.5-Flash from 49.5% to 68.1% and Qwen-3.6-27B from 39.4% to 63.3% across five benchmarks. Larger models benefited more, but smaller models using WikiSkill could match the performance of larger models without it.
What to watch
The framework's gains varied by task type, with the biggest improvements in math and spreadsheet tasks. Skills developed by one model often transferred to another, sometimes even working better than the receiving model's own skills, but transferability should be checked case by case.

Ask the AI about this article →

Summaries like this, in your inbox every morning.

Get news like this every morning — free

Context & Analysis

WikiSkill addresses a core limitation of AI agents: they don't truly learn continuously from each run. Instead, the framework has the agent write better instructions for itself after each task and store them in a persistent wiki, building up cumulative knowledge. This workaround, while not elegant, proved effective in the study's tests.

The architecture separates experience (raw traces), knowledge (the wiki), and action (skills). A Wiki Maintainer distills patterns from execution traces, a Skill Proposer suggests changes, and a gating mechanism validates them on a separate set before deployment. If a change fails, it's rolled back, but the wiki keeps the failed attempt for future reference.

Notably, skills transferred between models, sometimes outperforming the receiving model's own skills. However, the researchers advise checking transferability case by case, as it isn't always beneficial. The approach draws on Andrej Karpathy's concept of an "LLM Wiki," applying it to automatic skill development.

FAQ

What are the three layers in WikiSkill?: The Raw Layer stores full execution traces, the Wiki Layer holds distilled insights that never reset, and the Skill Layer contains active instructions that can be rolled back if they hurt performance.
Does WikiSkill work for small models?: Smaller models like Qwen-3.5-4B struggled with long multi-step tasks, but they still improved. Small models using WikiSkill could match larger models without it.

THE DECODERRead Original Article

Send on LINE

Get the latest Large Language Models news every morning

For example, today's edition would include:

AI job losses so far: only 3% of US workersFortune AI · 1h ago
Chatbot on Your Own PC: Privacy, Offline, FreeWIRED AI · 1h ago
Meta open-sources Muse Glimmer, keeps Muse Spark 1.2 closedYahoo Finance AI · 4h ago

AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.

Subscribe free →

Free · takes 30 seconds · unsubscribe anytimeWhat is AIToday? →