Guo Daya has filled the last shortcoming at ByteDance (字节跳动)

Zhipu (智谱) gives its agent a memory — and a store

Zhipu (智谱) has pushed its AutoClaw (中文名“澳龙”) agent into a new phase: autonomous learning plus a Skills marketplace. The company announced that AutoClaw now supports a self‑evolution mechanism that scans each conversation for user corrections, new teaching methods, expression preferences and its own errors, then surfaces "evolution request" cards for the user to approve and commit to long‑term memory. The agent was already notable for one‑click local deployment as an OpenClaw smart agent; this update adds persistent personalization.

Under the new flow, a single user correction can become a lasting capability. Tell AutoClaw to "reply more concisely" and, after approval, it will shorten future replies automatically; make it follow a corrected process once, and similar tasks will thereafter take the right path. It has been reported that this kind of on‑device, user‑approved continual learning addresses a gap in personalization that large consumer platforms — ByteDance (字节跳动) among them — have wrestled with when scaling AI experiences across millions of users.

Skills: office automation and expert playbooks

Alongside self‑evolution, Zhipu launched a Skill store. The initial offerings are a GLM Office Skills five‑pack for PPT, DOCX, XLSX, PDF and Charts — essentially off‑the‑shelf, scenario‑tuned document generation for everyday office work. Zhipu also packaged expert methods into ready‑to‑use Skills: examples include a "Nuwa" Skill, the 横纵分析法 (horizontal‑vertical analysis) and a Frontend Slides workflow, lowering the barrier so ordinary users can "out‑of‑the‑box" apply specialist techniques.

This move is also strategic. As Chinese AI firms race to build more autonomous, privacy‑friendly agents — and as Western export controls and geopolitical tensions push domestic players toward self‑reliant stacks — startups are differentiating on personalization, local control and expert‑driven tooling. Will skills and memory be enough to tilt users away from incumbents? It has been reported that some observers see AutoClaw's combination as precisely the missing piece for broader consumer adoption.