Zhipu AutoClaw launches the "self-evolution" mechanism, where an agent's one-time mistake becomes permanent memory, simultaneously releasing the Skill Store.

ME News update, April 17 (UTC+8). According to Beating Monitoring, Zhipu AI’s Agent framework AutoClaw (AoLong) has officially launched a self-evolution mechanism. The mechanism is designed to address Agent “forgetfulness” by identifying user corrections, new approaches, or expression preferences in conversations, and turning them into the Agent’s permanent memory. Users no longer need to repeatedly remind it to “reply more concisely” or “don’t use dashes,” because after being taught once, the Agent can evolve on its own. The self-evolution mechanism is triggered by keywords (such as “from now on” or “remember”) or by automatic detection (refining/summarizing experience after complex tasks). After each round of conversation, AutoClaw scans the dialogue and pops up an “evolution request” card, which is only written into memory after the user approves it. To avoid information noise, Zhipu has set high-quality evolution principles, favoring 1-3 in-depth evolutions per week rather than daily massive fragment record-keeping. This mechanism helps the Agent shift from “passively executing instructions” to “continuously growing through use”—the more users teach it, the more closely it matches the user’s personal habits.

Also launched at the same time is the AutoClaw Skill Store, featuring the first batch of Zhipu’s self-developed GLM Office Skills suite (PPT, DOCX, XLSX, PDF, Charts). Powered by the GLM-5.1 model optimized for office scenarios, it supports intelligent self-check formatting and layout, cross-format conversion (such as Word to PDF), and synchronized generation across multiple files. In addition, the store introduces expert co-created Skills, including the “Nüwa Skill,” which can replicate character thinking, and the “Horizontal and Vertical Analysis Method” for in-depth research—packaging top industry methodologies into ready-to-use Agent capabilities. (Source: BlockBeats)