10xAI Wiki

15分钟阅读

Knowledge Management in the AI Era

The Central Problem

How do you maintain a body of knowledge that AI can use effectively over time — not just for a single session, but persistently, with cross-references, synthesis, and cumulative learning?

Traditional note-taking assumes human readers. AI-native knowledge management assumes AI readers — a fundamentally different design challenge.

The LLM Wiki Pattern

This vault uses the LLM Wiki pattern, inspired by Andrej Karpathy’s work on LLM-based personal knowledge systems.

Core Architecture

Raw Sources (raw/)          → Immutable, AI reads but never modifies
    ↓ Ingest
Wiki (wiki/)                → LLM-written and maintained
    ├── sources/            → Source summaries
    ├── entities/            → People, companies, concepts
    ├── topics/              → Synthesized domain knowledge
    ├── analyses/            → Custom comparisons, insights
    └── index.md + log.md    → Schema and tracking

Why This Beats Traditional RAG

RAG (Retrieval-Augmented Generation): On each query, retrieve relevant documents from a corpus and add to context.

Problems with RAG:

  • No cross-references between documents
  • Retrieved documents may conflict with each other
  • No synthesis — just retrieval
  • Same information retrieved multiple times in different forms

LLM Wiki advantages:

  • Cross-references already exist in the wiki
  • Contradictions are flagged and resolved
  • Synthesis is cumulative — built once, maintained incrementally
  • Knowledge is compiled for reuse, not re-derived on each query

Karpathy’s position: RAG is dead. The context window problem it solves will disappear as windows grow. The real solution is a compiled, maintained knowledge artifact.

Karpathy’s Method — Six Steps

From “AI大神卡帕西的知识管理方法刷屏，用Get笔记六步抄作业”:

  1. Capture everything — Save articles, notes, screenshots without judgment
  2. Use AI to process — Let AI read and extract key insights
  3. Link and cross-reference — Connect new information to existing knowledge
  4. Review regularly — AI-generated summaries keep knowledge fresh
  5. Use for queries — The wiki becomes the knowledge source for AI
  6. Iterate and improve — Each interaction refines the knowledge base

Key Tools in His Stack:

  • nanochat — Minimal ~500-line ChatGPT implementation as teaching tool
  • nano RAM — Minimal working memory demonstration
  • Personal OS for AI agents — File system as database

Obsidian as the Foundation

Obsidian is the preferred tool for LLM Wiki implementations because:

FeatureBenefit
Local Markdown filesNo lock-in, git-versionable
Wiki-linksNative cross-referencing
Dataview pluginFrontmatter queries for dynamic organization
Graph viewVisualize knowledge connections
Web ClipperSave web content directly
TemplatesConsistent note structure

Key Obsidian Plugins for AI Knowledge Management:

PluginPurpose
DataviewQuery notes by frontmatter tags/dates
Web ClipperSave articles with one click
TemplaterStructured note templates
Graph ViewVisualize connections
Local REST APIExpose vault to AI agents

Graphify — Knowledge Visualization

Graphify (71.5x improvement claim) converts Obsidian notes into knowledge maps:

  • Parses all vault notes
  • Identifies entities, concepts, relationships
  • Generates visual knowledge graphs
  • Highlights bridge concepts (connecting different domains)
  • Makes implicit relationships explicit

Use case: When starting a new project, use Graphify to see all existing related notes — surfacing connections humans might miss.

Context Engineering — Designing AI’s Information Diet

The most sophisticated knowledge management systems are designed with context engineering in mind:

The principle: LLMs don’t treat all context equally. Information earlier in the context gets more “attention weight.” System prompts (like SOUL.md, AGENTS.md) should be written to be read first and carry the most weight.

Structured context hierarchy (from OpenClaw):

  1. Identity (SOUL.md) — Who is the agent? What are its values?
  2. Operations (AGENTS.md) — What rules govern its work?
  3. Skills (loaded on demand) — What specific capabilities does it need now?
  4. Shared context — What do multiple agents need to know?

The “Second Brain” Reborn

The “Building a Second Brain” (BASB) movement (Tiago Forte) predates the AI era but has been completely transformed by it:

BASB (Pre-AI)AI-Native
Notes for human future selfNotes for AI agents
Progressive summarizationAI synthesizes continuously
Organize by project/actionOrganize by entity/concept
Retrieve when neededInject proactively

The key shift: the second brain is no longer for human retrieval — it’s for AI context. This changes what makes a note “good”: not how useful it is when re-read, but how useful it is when injected into an AI’s context window.

Practical Implementation in This Vault

This vault follows the LLM Wiki schema in CLAUDE.md:

  • raw/: Immutable sources — never modified
  • wiki/: LLM-owned and maintained — the compiled knowledge
  • wiki/log.md: Activity tracking — what was ingested, when
  • wiki/index.md: Master navigation — all pages in one place
  • Frontmatter: YAML metadata for Dataview queries

Ingest workflow: New source → Read → Extract insights → Create/update wiki pages → Update index → Log

Key Tools & Systems

LLM Wiki Implementations

ToolDescriptionLink
Graphify知识图谱生成，71.5x token 节省Graphify
GBrainYC CEO 开源的全息记忆系统GBrain
LLM Wiki跨平台桌面应用（3300+ Stars）2026-04-29-LLM-Wiki-OpenSource
Wechat-Cli微信聊天记录导入 WikiWechat-Cli

Knowledge Management Tools

ToolDescriptionLink
NotebookLMGoogle 的知识利用工具NotebookLM
Cubox个人阅读记忆系统Cubox
Obsidian SkillsAI 智能体精通 Obsidian2026-03-08-Obsidian-Skills-AI-Knowledge-Management

Second Brain Systems

SystemCore ConceptSource
Karpathy V2.0三步编译法 + 复利效应2026-04-07-Karpathy-LLM-Wiki-Second-Brain
Get笔记六步法对话式实现 Karpathy 方法2026-04-07-Karpathy-GetNote-Six-Steps
GBrain三层记忆 + 全息记忆2026-04-13-YC-CEO-GBrain-OpenSource

三层架构核心概念

raw/ → wiki/ → outputs/

raw/                    # 原始资料，不做修改
  articles/             # Web Clipper 剪藏的文章
  podcasts/             # 播客转录
  tweets/              # 推文收藏
wiki/                  # 编译产物，由 LLM 维护
  summaries/           # 逐篇三步编译摘要
  concepts/            # 概念条目
  indexes/             # 自动维护的索引
outputs/                # 运行时输出
  qa/                  # 问答沉淀
  health/              # 健康检查报告

三步编译法

  1. 浓缩（第一性原理 + 剃刀法则）：抓主线，用剃刀法则删减
  2. 质疑（芒格反驳法）：逻辑链检查、前提假设、不适用场景
  3. 对标（它山之石可以攻玉）：跨域类比、迁移应用

AI Deep Research Methodology

Anthropic 的 Research（深度研究）模式展示了 AI 如何进行系统性知识调研：

  1. 任务拆解：将宽泛问题分解为研究矩阵（层级 x 内容类型 x 行业 x 语言）
  2. 广度搜索：大量短关键词建立全景认知（找到地图）
  3. 深度搜索：围绕关键节点深入展开（案例深挖）
  4. 交叉验证：同一数据点 2-3 来源确认
  5. 寻找原始来源：从二手报道追溯到一手资料（X 原帖、学术论文 PDF）
  6. 填补空白：发现维度缺失后补搜
  7. 综合与结构化：上百数据点整合为有内在逻辑的叙事

信息优先级：一手 > 专业垂直 > 主流媒体 > 二手分析。主动寻找批判性反叙事平衡 VC 乐观叙事。

详见 ../sources/深度搜索如何实现。这对 LLM Wiki 的 Query 工作流有直接参考价值。

Enterprise Deep Research Backend Pattern (NVIDIA AI-Q)

NVIDIA AI-Q (source) turns deep research into a deployed knowledge pipeline rather than an improvised prompt workflow. Compared with generic web research, AI-Q emphasizes enterprise constraints:

  1. Data stays where it lives: AI-Q can run inside the same governed environment as internal documents.
  2. Auth-aware retrieval: MCP servers can be connected with service accounts or signed-in user tokens.
  3. Clarification before retrieval: a human-in-the-loop clarifier resolves ambiguity before long-running research begins.
  4. Cited reports as output: the harness receives structured, source-attributed synthesis rather than raw document access.
  5. Evaluated research quality: FreshQA, Deep Research Bench, DeepSearchQA, and local eval harnesses make research quality measurable.

For an LLM Wiki, the lesson is that “deep research” should be treated as a workflow with routing, source attribution, evaluation, and audit logs—not merely as a longer answer.

Sources

AIHOT — Information Curation System (2026-05)

From AIHOT: A practical example of AI-assisted knowledge management for content creation. Monitors 168 sources, uses DeepSeek V3.2 for pre-screening and V4 Pro for 5-dimension scoring. Key lesson: iterated scoring 11 times, ultimately separating model scoring from code-based decision making. Model only does scoring; weighting, thresholds, and curation handled by deterministic code. Prompt reduced from 600 lines to 200 lines.

Academic Writing as Knowledge Management (2026-05)

From nature-skills: Packaging Nature-standard writing rules into AI skills (figure generation, polishing, citation search, paper-to-PPT, data compliance). Core philosophy: “Scientists’ time should be spent on science” — automates the “unscientific but important” manual work.

研究助理型知识库与私有数据路线之争 [2026-08-06]

  • 研究助理型知识库：Claude Code 9h 知识库——Claude Code 长会话搭建 Obsidian 四层库（papers/concepts/ideas/surveys），每日自动”选文→精读→推送→反馈→回写”闭环；“主干节点”（高链接概念）成为生长中心。与本 vault LLM Wiki 高度同构，但缺不可变 raw 层（溯源/防幻觉弱）。
  • 内容生产导向（桑梓三库法）：经历/成绩/风格三库，查询时上传文件的无状态喂料——与编译型路线对照，见 源页。
  • 私有数据 Deep Research 的路线之争：DeepSearcher（Zilliz，LLM + Milvus 向量检索，数据不出域）代表 RAG 路线；本页既有的 Karpathy “RAG 已死、编译型知识库” 立场为对立路线。私有数据深度研究最终走向向量检索还是知识编译，值得跟踪（星标数矛盾已在实体页标注）。

KnowFlow vs Linkly AI：学习流程原生化 vs 文献库 MCP 接入 [2026-08-10]

| 维度 | KnowFlow | Linkly AI | |------|-----------|------------| | 问题规模 | 单篇学习流程 | 大规模文献库（1000+ 篇论文） | | 流程 | 剪藏 → 清洗 → 摘要 → 知识图谱（Mermaid）→ 测验 | search → outline → read 渐进式消费 | | 接入方式 | Obsidian 原生插件（侧边栏随当前笔记切换四种上下文模式） | MCP Server（暴露给 Claude、Cursor 等任意 Agent 客户端） | | 产物落点 | Markdown/frontmatter（本地、弃用无锁定） | 本地文件原话引用（反幻觉） | | 成熟度 | 0.1.0 闭源，承诺 1.0.0 开源 | 产品方自述配置流程，无第三方基准 |

两者代表 AI 时代知识管理的两种互补形态：单篇深度学习原生长进笔记环境（把 AI 能力搬进数据所在地），大规模语料服务化供任意 Agent 消费（把数据经标准协议暴露出去）。共同出发点是消灭工具切换与手工上下文传递的摩擦——KnowFlow 作者从 Codex Skills 迁移回 Obsidian 插件，正是因为”每处理一篇文章都要离开 Obsidian 重新描述文件”。对照分析见 AI 工作流的原生化模式。

来源：KnowFlow、Linkly AI（两者均为产品自述，待第三方验证）。

关系图谱

反向链接