Agentic LLM Wiki Template

A forkable agentic LLM Wiki template for turning sources into maintained, interlinked markdown knowledge.

English | 繁體中文

What Is This?

Most document workflows use retrieval as the end state: upload files, search chunks, answer a question, then repeat the same rediscovery next time. This template takes a different stance.

An LLM Wiki is a persistent markdown knowledge base that agents help maintain. Raw sources are captured once, then important information is compiled into durable pages, cross-linked, queried, corrected, and promoted when it becomes worth remembering.

The human role is to curate sources, ask useful questions, and review judgment. The agent role is to summarize, route, cross-link, maintain indexes, preserve citations, and make write-back decisions explicit.

Why This Exists

RAG, file uploads, and scattered notes can answer questions, but they do not automatically accumulate understanding. A useful knowledge base needs more than search:

• source-backed capture
• maintained topic and object pages
• explicit memory promotion
• visible contradictions and stale claims
• agent instructions that make future sessions behave consistently

This repository is a template for that operating model. It is not a desktop app, hosted service, or full vector database. It is a markdown repo designed to be forked, customized, and operated by humans plus coding agents such as Codex, Claude Code, Gemini CLI, or similar tools.

Architecture

agentic-llm-wiki/
  00_Control-Tower/        navigation, tasks, decisions, risks, activity logs
  01_Clients/              account-level truth for each client or domain
  02_Projects/             delivery/project truth, meetings, deliverables
  03_Product/              product or system source of truth
  04_Knowledge/            reusable concepts, playbooks, and patterns
  05_AI_System/            AI/agent/RAG/automation runbooks
  06_Research/             source-facing external research notes
  07_Sources/              raw evidence and source capture
  08_Templates/            reusable note templates
  09_Memory/               promoted durable facts, decisions, context packets
  90_System/               governance, schemas, routing rules, agent rules
  .skills/                 workflow wrappers for agents
  scripts/                 small local helper commands

The core flow is:

07_Sources -> canonical wiki -> 09_Memory -> 90_System governance

How It Works

1. Capture: place a transcript, article, meeting note, or source summary under 07_Sources/.
2. Ingest: run a helper command or ask an agent to create a source note with metadata.
3. Write back: update the maintained page that should now be more correct.
4. Query: answer from canonical pages first, then memory, then raw sources when needed.
5. Lint: periodically check contradictions, stale claims, missing pages, weak links, and promotion gaps.

Every query should end with a write-back decision: stay raw only, update canonical wiki, or promote to memory.

Quick Start

git clone https://github.com/ericweichun/agentic-llm-wiki.git
cd agentic-llm-wiki
./scripts/wiki-client-init ACME --name "Acme Corp"
printf 'Synthetic kickoff note for ACME. Scope is still tentative. Main risk is unclear ownership.
' > /tmp/acme-kickoff.txt
./scripts/wiki ingest client/ACME "Kickoff note" /tmp/acme-kickoff.txt --meeting
./scripts/wiki query "What do we know about ACME?" --domain client/ACME

The included ACME, Atlas, and Northstar notes are synthetic examples. Replace them after you fork the repository.

Agent Entrypoints

• Codex / OpenAI agents: read AGENTS.md, then .skills/company-orientation/SKILL.md.
• Claude Code: read CLAUDE.md, then .skills/company-orientation/SKILL.md.
• Gemini: read GEMINI.md, then .skills/company-orientation/SKILL.md.

Example Walkthrough

Synthetic ACME flow:

1. Add source material to 07_Sources/clients/ACME/.
2. Update 01_Clients/ACME/ACME - Client Home.md if account truth changed.
3. Update 02_Projects/ACME/ACME - Project Home.md if delivery truth changed.
4. Promote only durable cross-session context into 09_Memory/Clients/ACME/ACME Context Packet.md.
5. Keep one-off or unverified details in the source note.

Synthetic Atlas flow:

1. Product source evidence starts under 07_Sources/core/products/Atlas/.
2. Maintained product truth lives under 03_Product/Atlas/.
3. Reusable implementation lessons move to 04_Knowledge/ or 05_AI_System/ only after sanitization.

Fork Guide

1. Fork this repository.
2. Make your fork private before adding real sources or confidential notes.
3. Replace synthetic ACME, Atlas, and Northstar examples with your own domain objects.
4. Update 90_System/Company/ governance to match your naming, classification, and promotion rules.
5. Keep raw confidential sources out of public Git history.
6. Treat retrieval infrastructure as a helper, not the source of truth.

---

繁體中文

一個可 fork 的 agentic LLM Wiki 模板，用來把原始來源轉成可維護、可互相連結的 Markdown 知識庫。

這是什麼？

多數文件型 AI 工作流把檢索當成終點：上傳檔案、搜尋片段、回答問題，下次再重新找一次。這個模板採用不同假設。

LLM Wiki 是一個持續存在的 Markdown 知識庫，由 agent 協助維護。原始來源先被保存，重要資訊再被編譯到穩定頁面、建立連結、被查詢、被修正，必要時提升成 durable memory。

人的工作是選來源、問問題、審查判斷；agent 的工作是摘要、分流、交叉連結、維護索引、保留引用，並把是否寫回知識庫的決策講清楚。

為什麼需要它？

RAG、檔案上傳和零散筆記可以回答問題，但不一定會累積理解。一個真的能複利的知識庫需要：

• 有來源依據的 capture
• 持續維護的主題頁與物件頁
• 明確的 memory promotion
• 看得見的矛盾、過期 claim、缺口
• 讓未來 agent session 能一致工作的規則

這個 repo 是這套操作方式的模板。它不是桌面 app、SaaS，也不是完整向量資料庫；它是一個可以 fork、客製化，並讓人類與 coding agent 一起操作的 Markdown repo。

架構

核心流程：

07_Sources -> canonical wiki -> 09_Memory -> 90_System governance

• 07_Sources/ 保存 evidence 和 provenance。
• canonical wiki 位於 01_Clients/, 02_Projects/, 03_Product/, 04_Knowledge/, 05_AI_System/。
• 09_Memory/ 只放經過提升的 durable facts、decisions、context packets。
• 90_System/ 告訴 agent 如何安全地維護這個 wiki。

如何運作

1. Capture：把 transcript、文章、會議紀錄或來源摘要放進 07_Sources/。
2. Ingest：用 helper command 或 agent 建立帶 metadata 的 source note。
3. Write back：更新那個現在應該更正確的 maintained page。
4. Query：先查 canonical pages，再查 memory，必要時才回 raw sources。
5. Lint：定期檢查矛盾、過期資訊、缺頁、弱連結和 promotion gaps。

每次 query 都要結束於明確 write-back decision：stay raw only、update canonical wiki 或 promote to memory。

快速開始

git clone https://github.com/ericweichun/agentic-llm-wiki.git
cd agentic-llm-wiki
./scripts/wiki-client-init ACME --name "Acme Corp"
printf 'Synthetic kickoff note for ACME. Scope is still tentative. Main risk is unclear ownership.
' > /tmp/acme-kickoff.txt
./scripts/wiki ingest client/ACME "Kickoff note" /tmp/acme-kickoff.txt --meeting
./scripts/wiki query "What do we know about ACME?" --domain client/ACME

內建的 ACME、Atlas、Northstar 都是假資料範例。fork 之後請換成你自己的 domain。

Agent 入口

• Codex / OpenAI agents：讀 AGENTS.md，再讀 .skills/company-orientation/SKILL.md。
• Claude Code：讀 CLAUDE.md，再讀 .skills/company-orientation/SKILL.md。
• Gemini：讀 GEMINI.md，再讀 .skills/company-orientation/SKILL.md。

Fork 指南

1. Fork 這個 repo。
2. 加入真實來源或機密筆記前，先把 fork 設成 private。
3. 把 ACME、Atlas、Northstar 假資料換成自己的 domain objects。
4. 修改 90_System/Company/ 讓治理規則符合你的命名、分類、promotion 流程。
5. 不要把 raw confidential sources 放進 public Git history。
6. 把 retrieval infrastructure 當加速器，不要當真相來源。