Google gives AI agents a memory of past work
OpenAI opens an experiment for agent-ready websites, while a London trial takes AI into brain surgery.
By Saeed Ezzati7 min read
The audio edition
Listen to this newsletter
0:003:42
Read transcript
Google Research’s WikiSkill points to a different way for AI agents to improve: not retraining the model, but preserving what happened during work and testing better instructions for next time. In trials across five benchmarks, Gemini 3.5 Flash’s average score rose from 49.5 percent to 68.1 percent when it used the evolving skills. The framework separates experience from action. A Raw Layer stores complete traces, including tool calls and results. A persistent Wiki Layer turns those records into documented failures and successful tactics. Then a Skill Layer holds the instructions used on new tasks. A maintainer distills the evidence, a proposer suggests changes, and a separate validation set decides whether those changes stay active. If performance falls, the prior skill can be restored without deleting the experience behind the failed attempt. The gains were especially large in math and spreadsheets. Gemini rose from 33.0 to 72.6 percent on LiveMath, and from 50.5 to 76.6 percent on SpreadSheet. Qwen-3.6-27B’s five-benchmark average climbed from 39.4 to 63.3 percent. Improvements were smaller on OfficeQA, a long-document question-answering task. The caveat matters. Larger models generally benefited more, while smaller ones could lose multi-step search strategies in long contexts and revert to default behavior. Skills sometimes transferred between models, but not consistently. WikiSkill is therefore less a universal memory upgrade than a controlled process for turning operational evidence into candidate procedures, then checking whether they actually work on the model and task at hand. That emphasis on controlled deployment also appears in OpenAI’s ten-day WebMCP Challenge. WebMCP is an experimental standard for letting websites publish structured actions that agents can call directly, instead of forcing them to guess through a visual interface. The tools run in the webpage’s JavaScript context and share the user’s session. For now, the scope is narrower than the broader Model Context Protocol: named functions, descriptions, and typed inputs. Chrome 146 has a developer trial, while other major browsers had not shipped implementations. The challenge tests developer interest, not a settled web standard. The same question—where to put the controls—drives Bill Gates’s proposal to tax robots and AI tokens, and create a “Human Reserved” category for work where AI would be restricted or barred. Gates says the revenue could fund retraining and social programs while slowing displacement, including in sensitive encounters such as telling a patient they have an incurable disease. The International Federation of Robotics argues that taxing production tools could weaken investment, productivity, and competitiveness. Gates has not specified who would set the rules, which technologies qualify, or which jobs would be protected. And in London, the control boundary was visible in a clinical trial. Neurosurgeons used live AI video analysis while removing Rhys Hibbert’s 11-millimeter pituitary tumour. The system color-coded the gland, nerves, blood vessels, instruments, and tissue interactions, but it did not make surgical decisions; the team retained control. Officials called it the world’s first successful AI-assisted operation of this kind, and Hibbert’s recovery was positive, but one patient cannot establish a safety advantage. Across these stories, the practical frontier is clear: useful AI is moving into memory, interfaces, policy, and medicine—but each deployment still needs an explicit validation gate, a defined human role, and evidence that transfers beyond the original case.
