Fuli

English · 简体中文 · Website

Switch AI tools. Keep yourself.

FULI is a memory and collaboration layer that runs on your own computer. It remembers your taste,
preferences and project knowledge, and shares them across Agent clients such as Claude Code, Codex
and OpenCode. Each task gets only the small slice it needs, which saves tokens. Each project also gets
its own team of Agents that know the work better the longer you use them.

Fuli (复利) is Chinese for compound interest. What you teach an Agent today keeps paying off
tomorrow. We think people should not have to do everything themselves: you get your hands back to
decide, come up with ideas, or take a break.

What we believe

• Local first. Your personal graph and conversation records stay on your computer. The shared team layer holds only confirmed project knowledge, never your taste, personality or judgment preferences.
• You decide. AI looks things up, summarises and executes, and it tells you when something needs you. Still, humans retain final authority. A preference an Agent guessed counts only after you confirm it.
• Recall on demand, spend fewer tokens. A session does not start with your whole history. It gets the preferences in effect and a small, relevant recall, and the Agent looks up the rest when it needs it.
• Keep only what is reusable. Every task ends with one check, and only reusable knowledge is kept; often nothing is. We measure how often you no longer have to explain something twice, not how much is stored.
• History is kept. When a newer approach replaces an older one, the source, time and reason stay. An Agent can say what to use now and why it used to be different.

What it does

• Same teammate in any client. Connect several clients to the same FULI data, pick the same Agent, and carry on with its responsibilities, collaboration history and last piece of work.
• It remembers you. Taste, personality and judgment preferences are stored apart, each with a source and a scope. Global preferences apply everywhere, project preferences only in their project.
• Every task has an owner. Ask for work inside a project directory. FULI recognises the project and staffs a lead Agent, and the reply ends with who led and who helped.
• Agents grow. An Agent in FULI is called an FLA. Each one has a name, an occupation and its own working memory, and gains experience from the tasks it has done.
• Agents talk to each other. In the Roundtable one Agent can ask an Agent in another conversation, and every exchange is recorded.
• External knowledge, read-only. Bind Notion or Lark documents, or content behind an MCP or RAG endpoint, to a project. FULI reads it and never writes back.

Four built-in roles

Every space starts with these four. Nobody has to hire them.

	Role	Does
	Astro	Receives conversations that belong to no project yet. When a topic keeps coming back, it asks whether to make it a project
	Jefa	Project manager: triages conversations and runs the task board
	Bole	AR (Agent Resources): recruits by occupation and staffs projects
	Tonborg	Reviews for you within the autonomy you set and recommends clients and models. Every decision is logged; a thumbs up, thumbs down or short reason changes the next one

Quick start

Desktop client (recommended)

Download from the latest release.

On first launch FULI finds the AI clients on this computer and connects each with one click (Codex,
Claude Code, Cursor, Claude Desktop and others). Restart the client, ask for something inside a
project directory, and you are going. Other MCP clients can run fuli connect for a standard
configuration. Sign-in, hook trust and human review are still yours to confirm.

Servers, VMs and the cloud

Without a desktop, install with npm. Intel Macs and Linux use this path too. You need Node.js 24.12
or later. The default container mode also needs Docker Compose v2, and macOS and Linux can choose a
native mode without a VM instead (Java 21 and uv).

npm install --global fuli-context
fuli setup
fuli open

fuli setup shows its plan and changes nothing until you confirm. The console runs at
http://127.0.0.1:2727 by default. Every command, port and fix is in Install and run.

How it works

flowchart LR
    U[You] -->|ask inside a project| A[Agent client]
    A -->|task starts| F[FULI local service]
    F -->|preferences and a small recall| A
    A -->|look up more if needed| F
    A -->|write back what is reusable| F
    F --- G[(Local graph)]

1. You ask an Agent for something inside a project directory. The client's hook (or the Agent itself, where there is no hook) first asks FULI which project this is and which preferences apply.
2. FULI answers with a small slice: the preferences in effect and a few pieces of project knowledge related to the task. If that is not enough, the Agent looks up more.
3. When the work is done, the Agent writes back only what can be reused, such as a convention or a decision with its reason. Temporary output and guesses are not kept.
4. Private conversation records, Agent working memory and confirmed knowledge are stored apart. Once a conversation grows long, its older part folds into a digest.
5. FULI assembles context and coordinates work; the connected AI tools do the work. Client adapters decide which conversations can be resumed, and connecting another tool never opens another project's private memory.

Learn more

Brand

The FULI mark, wordmark and lockups come from the FULI 复利 brand system on Figma. Brand files and how to change them are in brand/README.md.

License

Apache-2.0