Google WikiSkill lifts Gemini-3.5-Flash agent scores from 49.5% to 68.1%
Google Research researchers unveiled WikiSkill, an architecture that equips agents with durable, wiki-structured stores for lessons from prior attempts. Rather than resetting context after every execution, agents can reference documented errors and wins when tackling new work. Average reported scores climb sharply: Gemini-3.5-Flash moves from 49.5% to 68.1%, and Qwen-3.6-27B from 39.4% to 63.3%. Independent coverage highlights that bigger base models extract greater value, while compact models armed with WikiSkill may still close part of the gap. In the same news cycle, Reuters-via-Techmeme reported Google pushing Gemini Enterprise for Legal with specialized agents wired to Thomson Reuters, LexisNexis, and Harvey, illustrating how Google is pairing agent tooling with domain integrations. Available excerpts stop short of complete evaluation protocols.
Google WikiSkill lifts Gemini-3.5-Flash agent scores from 49.5% to 68.1%
Researchers at Google Research have introduced WikiSkill, a framework that pairs AI agents with a persistent knowledge base. On average, the framework boosts Gemini-3.5-Flash from 49.5 percent to 68.1 percent and Qwen-3.6-27B from 39.4 percent to 63.3 percent.
Key takeaway
WikiSkill indicates agents can compound performance by retaining documented mistakes and successes in a persistent knowledge base across runs.
What happened
Researchers at Google Research introduced WikiSkill, a framework that pairs AI agents with a persistent knowledge base, according to The Decoder. Agents document both failures and successes in a wiki-like structure instead of discarding what they learned after each run.
The Decoder reports that WikiSkill lifted average scores for Gemini-3.5-Flash from 49.5 percent to 68.1 percent and Qwen-3.6-27B from 39.4 percent to 63.3 percent. Coverage also notes larger models benefit more, while smaller models with WikiSkill can match stronger performance.
Evidence
Google Research introduced WikiSkill as a framework giving agents a persistent knowledge base.
The Decoder · attributed
Google Research has introduced WikiSkill, a framework that gives AI agents a persistent knowledge base.
Agents record failures and successes in a wiki-like structure rather than discarding lessons after each run.
The Decoder · attributed
Instead of discarding what they learned after each run, agents document both failures and successes in a wiki-like structure and use that knowledge to get better over time.
WikiSkill raised Gemini-3.5-Flash average scores from 49.5% to 68.1%.
The Decoder · attributed
On average, the framework boosts Gemini-3.5-Flash from 49.5 percent to 68.1 percent
WikiSkill raised Qwen-3.6-27B average scores from 39.4% to 63.3%.
The Decoder · attributed
Qwen-3.6-27B from 39.4 percent to 63.3 percent.
Larger models benefit more from WikiSkill, though smaller models can still gain substantially.
The Decoder · attributed
Larger models benefit more, but smaller models with WikiSkill can match t
Google expanded Gemini Enterprise for Legal with specialized agents and integrations with Thomson Reuters, LexisNexis, and Harvey.
Techmeme · attributed
Google launches Gemini Enterprise for Legal, expanding its platform with specialized AI agents and integrations with Thomson Reuters, LexisNexis, and Harvey
Why it matters
For builders, the reported benchmark jumps suggest structured cross-run memory may be a practical lever before upgrading to heavier foundation models on agent workflows.
Limits and uncertainties
The Decoder excerpt on smaller models matching stronger performance is truncated mid-sentence.
The packet reports average benchmark percentages but not task definitions, datasets, or full WikiSkill methodology.
Coverage of Gemini Enterprise for Legal does not clarify how Gemini differentiates from Harvey or incumbent Thomson Reuters and LexisNexis tools.
Practical implications
Builders running episodic agents may pilot wiki-like knowledge stores to persist failure and success notes between runs.
Operators comparing model upgrades should benchmark whether memory frameworks like WikiSkill close gaps for smaller models on their workflows.
What to watch
Follow-on publications or Google releases detailing WikiSkill evaluation tasks, methodology, and the truncated smaller-model matching claim.
Signs of WikiSkill-style persistent memory being productized inside Gemini agent or enterprise legal offerings.
