2 stories in this blend

Prevent wrong conclusions from corrupting future task runs by reviewing proposed agent memories with read only source access. Following this process keeps historical agent data reliable and lowers processing costs.

Computer science researchers from Google and Purdue University developed a technique called SKILL.state that tracks an agent's current working memory instead of re-sending its full chat history. In testing, the approach reduced token consumption by 94 percent while slightly improving overall task accuracy.