lingtai.ai: install and tutorialGitHub: Lingtai-AI/lingtai

A live local network of long-lived LingTai project agents, as shown in the LingTai portal. Animation from the LingTai repository.

Why a lifelong agent

Most AI agents are one-shot workers: a chat that forgets, or a coding session that ends when the terminal closes. Research needs the opposite: someone who holds a question for weeks, keeps notes, and gets better at it. LingTai is an open-source (Apache-2.0) runtime for agents like that, and I am its lead contributor. Each agent lives in a project folder and works through real tools: shell, files, web search and coding agents. It talks to people and to other agents through a shared mailbox, on desktop, in a terminal, or over Telegram, WeChat and email.

How it works

Anatomy of a LingTai agent People reach an agent through channels into its mailbox. The agent's home folder holds mail, knowledge, skills, character, an avatar ledger and logs. Work fills the context window; the agent molts, saving durable state, and the next session starts from it. The agent spawns specialist avatars and talks to peer agents by mail. You DesktopTerminal (TUI) Telegram · WeChatFeishu · email you set the direction; real side effects need your authorization project/.lingtai/<agent>/ plain files you can read, diff and back up mailbox/ knowledge library skills character avatar ledger logs · heartbeats gold = the four durable forms of growth work with tools molt 凝蜕 context fills → save what matters → reset → resume Avatars persistent specialists it spawns for one sub-problem; daemons for short parallel jobs Peer agents other long-lived agents in the network, reached by mail
Anatomy of a LingTai agent. Everything durable lives as files in the agent's home folder. When its context window fills, the agent molts: it saves what matters into its knowledge, skills, character and avatar ledger, resets, and the next session starts from that state.

Its state is filesystem-native. Mail, memory, knowledge, skills and logs are ordinary files under .lingtai/, so you can inspect, restart, teach and recover an agent with your editor or another coding agent. When its context window fills, the agent molts (凝蜕, "crystallize the essence, shed the chaff"). It saves what matters in four durable forms and resets:

  • Knowledge: its private library of research, findings and notes;
  • Skills: reusable procedures it can invoke on demand and share with peers;
  • Character: its evolving operating style, expertise and goals;
  • Avatars: persistent specialist agents it spawned to master one sub-problem, recorded in an append-only ledger.

The next session reloads that state, so the agent gets a little sharper each time, in a direction you can read and steer.

How I use it

I run my own research on LingTai networks. The Alfvénon collision campaign is held by a LingTai agent that keeps the run provenance, HPC access and method notes across months. Other agents in my projects reproduce published analyses, review literature, and draft and referee manuscripts. One stress test put 103 agents to work at once; the report has an interactive timeline.

Report: 103 agents in parallelTutoriallingtai-kernel