
It chats inside Feishu, DingTalk...
When you host a similar agent yourself, skills often freeze on the day you install them, and a new machine means starting over. MaxHermes puts the accumulation into use. A skill document is extracted after a complex task and loaded again later. Cross-session memory keeps your feedback instead of opening a blank chat every time. Schedules are written in everyday language, and several sub-agents can run together. Usage draws on an existing credit plan, and the announcement says there is no separate API fee. It fits people who want a work-chat assistant that revises its own methods. It is not yet the product that already connects to a public skill community.
MaxHermes from MiniMax hosts a Hermes agent in a cloud sandbox. You do not bring a server or paste an API key. It chats inside Feishu, DingTalk, and WeCom and stays online. After a complex task it writes reusable steps into a separate skill document, loads that skill later, and revises it from new feedback. The same announcement describes cross-session memory, schedules written in plain language, and parallel sub-agents. Connecting a skill community and migrating a persona from an older assistant are listed as later updates.
Difficulty: Beginner Friendly
Skills after tasks
After a complex task, reusable steps are saved as a skill document, loaded when needed, and revised from later feedback.
Cross-session memory
The announcement says it remembers each piece of feedback and carries it into the next conversation.
Chat-app presence
Feishu, DingTalk, and WeCom can host the chat, so you do not open another console or leave your own computer on.
Schedules and sub-agents
Schedules are defined in plain language, and several sub-agents can run in parallel on parts of a complex job.
Ongoing tasks in Feishu
Teammates send work from Feishu, DingTalk, or WeCom so the cloud assistant stays online and keeps the last correction.
Turn a repeat flow into a skill
After a complex flow works once, it is saved as a skill document and loaded for the next similar task, then revised.
The announcement presents a hosted assistant in a cloud sandbox. You do not prepare a server or paste an API key. The chat entry is Feishu, DingTalk, and WeCom. Skills are extracted after complex tasks instead of being frozen at install time. Self-hosting means maintaining the environment and uptime. A skill community is still a later plan.
The announcement describes two layers. Cross-session memory keeps feedback for the next conversation. The learning loop writes a skill document after a complex task, loads it later, and revises it from new feedback. Schedules and parallel sub-agents are separate and do not replace those two layers.
Real reviews and feedback from users