Google's WikiSkill gives AI agents a persistent memory of past mistakes to sharpen future performance
In shortGoogle Research has introduced WikiSkill, a framework that gives AI agents a persistent knowledge base. Table: WikiSkill (highlighted) achieves the best or statistically equivalent performance across most model-benchmark combinations.
What happened
On individual benchmarks, the jumps can be bigger: Gemini-3.5-Flash climbs from 33.0 percent to 72.6 percent on LiveMath and from 50.5 percent to 76.6 percent on SpreadSheet.
Why it matters
Google's WikiSkill changes how agents store, retrieve, and reuse knowledge across repeated tasks. Teams should test retrieval quality, rollback behavior, and memory growth before depending on it in production.
Who it affects
Developers integrating Google's WikiSkill · Researchers evaluating Google's WikiSkill
The bigger picture
Google's WikiSkill moves the competitive test for agents beyond one-shot answers to whether stored knowledge can be corrected, governed, and reused across repeated work.
What happens next
- Watch independent tests of retrieval quality, rollback behavior, memory growth, and failure recovery for Google's WikiSkill.
This information is still under review.