LLM Wiki — Personal Knowledge Base

English | 简体中文

A personal knowledge base continuously maintained with Obsidian + LLM Wiki. Inspired by Karpathy's LLM Wiki pattern.

Core idea: Knowledge is not derived from scratch on every query — it is compiled once and continuously accumulated. Humans curate materials and ask questions; the LLM handles all the tedious organization work — summarization, cross-referencing, archiving, and consistency maintenance.

How It Differs from Traditional RAG

	Traditional RAG	LLM Wiki
Knowledge Storage	Raw documents + vector index	Structured, interlinked Markdown pages
Query Method	Retrieve and assemble from scratch each time	Answer based on pre-compiled knowledge
Knowledge Accumulation	None — re-derived every time	Yes — every operation makes the wiki richer
Cross-references	None	Automatically maintained [[wikilink]] network
Contradiction Handling	Unaware	Proactively annotated with sources cited

Architecture

├── AGENTS.md          # Schema spec (OpenAI Codex and other general agents)
├── CLAUDE.md          # Schema spec (Claude Code adaptation)
├── GEMINI.md          # Schema spec (Gemini CLI adaptation)
├── skills/                # Agent Skills (detailed workflow rules, loaded on demand)
│   └── llm-wiki/          # LLM Wiki skill
│       ├── SKILL.md       # Four subcommands: ingest/query/lint/publish
│       └── references/    # Shared spec documents (7 *-spec.md files)
├── raw/               # Raw materials (human-curated, LLM read-only)
│   ├── articles/      # Web articles, blog posts
│   ├── papers/        # Academic papers, white papers
│   ├── docs/          # Official documentation excerpts
│   ├── transcripts/   # Meeting notes, talks, podcast notes
│   └── assets/        # Images, diagrams, data files
├── wiki/              # LLM compiled artifacts (LLM read-write, human read-only)
│   ├── index.md       # Global index
│   ├── log.md         # Operation log
│   ├── log-archive/   # Log archive (by month)
│   ├── lifecycle.md   # [Pluggable] Detailed lifecycle data
│   ├── entities/      # Entity pages (tools, frameworks, people)
│   ├── concepts/      # Concept pages (patterns, methodologies)
│   ├── summaries/     # Material summary pages
│   ├── comparisons/   # Comparative analysis pages
│   └── synthesis/     # Synthesis analysis pages
└── output/            # Polished output (LLM generates, human reviews and finalizes)
    ├── posts/         # Blog posts, articles
    ├── reports/       # Research reports, technical survey reports
    ├── slides/        # Presentations (Marp format)
    ├── tutorials/     # Tutorials, step-by-step guides
    └── newsletters/   # Weekly, monthly, knowledge briefings

Core Principles

Compilation over Retrieval

Traditional RAG retrieves from raw documents and assembles answers from scratch on every query — knowledge never accumulates. LLM Wiki takes the opposite approach: each time material is ingested, the LLM compiles knowledge into structured wiki pages, and subsequent queries are answered based on pre-compiled knowledge. Knowledge is compiled, linked, and continuously optimized — just like code.

Knowledge Lifecycle Management

Knowledge doesn't stay valid forever once written. This project introduces a complete lifecycle management mechanism:

• Confidence Scoring: Each page carries a confidence score (0.0–1.0), dynamically adjusted based on source count, access frequency, and time decay. Knowledge cross-confirmed by multiple sources is more trustworthy than single-source knowledge.
• Forgetting Curve: Inspired by the Ebbinghaus forgetting curve, knowledge that hasn't been accessed or reinforced over time naturally decays in confidence. Different types decay at different rates — conceptual knowledge decays slowly, comparative analysis decays quickly.
• State Transitions: active → stale → archived. Being queried or confirmed by new sources can restore active status.
• Supersession: When new information explicitly refutes old conclusions, superseded_by/supersedes bidirectional links enable version chain traceability. Old versions are preserved but marked as outdated.
• Demotion: When high-level synthesis is disproven, it is demoted to stale pending re-verification. Lower-level entities and concepts are unaffected.

Layered Consolidation (Four-Tier Memory Model)

Knowledge progressively promotes from lower to higher tiers. Each tier is more compressed, more reliable, and has a longer lifecycle:

┌─────────────────────────────────────────────────┐
│  Procedural    output/*                          │  Ready-to-use deliverables
├─────────────────────────────────────────────────┤
│  Semantic      wiki/comparisons/ + synthesis/    │  Cross-material insights
├─────────────────────────────────────────────────┤
│  Episodic      wiki/entities/ + concepts/        │  Structured knowledge
├─────────────────────────────────────────────────┤
│  Working       wiki/summaries/                   │  Single-material summaries
└─────────────────────────────────────────────────┘

Promotion criteria: An entity/concept in a summary is mentioned by 2+ materials → create entity/concept; 3+ entities/concepts form a pattern → create comparison/synthesis; synthesis confidence ≥ 0.85 → suggest publishing as a deliverable.

Progressive Disclosure

Detailed rules are distilled into Agent Skills (skills/llm-wiki/). AGENTS.md only retains a quick overview of core concepts. When executing specific operations, the LLM loads skill instructions and references on demand, avoiding heavy context consumption on every conversation.

Core Capabilities

Capability	Description
Ingest	Read materials from raw/, extract entities and concepts, create/update wiki pages, automatically maintain cross-references
Query	Answer based on compiled wiki knowledge, prioritize high-confidence pages, archive valuable answers
Lint	Contradiction detection, orphan pages, missing references, cross-reference integrity, lifecycle decay and state transitions
Publish	Produce blog posts, reports, slides, tutorials, briefings, and other standalone deliverables from wiki
Lifecycle Management	Confidence scoring, forgetting curve decay, state transitions, supersession/demotion, layered consolidation promotion
Pluggable Extension	Lifecycle is an enhancement layer — deleting lifecycle.md does not affect core wiki functionality

Workflow

All operations are executed through the llm-wiki skill (see skills/llm-wiki/SKILL.md for details):

Operation	Usage	Description
Ingest	/llm-wiki ingest <raw/path/file>	Read material → extract key points → create/update wiki pages → maintain cross-references → update lifecycle
Query	/llm-wiki query <question>	Answer based on compiled wiki knowledge, prioritize high-confidence pages, optionally save valuable answers back to wiki
Lint	/llm-wiki lint	Check for contradictions, orphan pages, missing references, lifecycle decay and state transitions — keep wiki healthy
Publish	/llm-wiki publish <type> [topic]	Produce polished output from wiki; LLM generates draft, human reviews and finalizes

Core Principles

• Schema Layer (AGENTS.md): Tells the LLM how to maintain the wiki — quick overview of core concepts
• Skills Layer (skills/llm-wiki/): Detailed workflow rules, loaded on demand, progressive disclosure
• Raw Layer (raw/): Immutable raw materials — the source of truth
• Wiki Layer (wiki/): LLM's compiled artifacts — structured representation of all knowledge
• Output Layer (output/): Deliverables distilled from wiki — standalone, audience-facing, independently readable
• Knowledge compound interest: Every operation makes the wiki richer
• Lifecycle management: Knowledge has temperature — frequently used knowledge warms up, long-dormant knowledge cools down
• Detailed specs in AGENTS.md, full workflow in skills/llm-wiki/SKILL.md

Quick Start

1. Choose Your LLM Agent

This project provides three equivalent schema files — pick the one matching your tool:

Tool	Schema File	Description
OpenAI Codex / General	AGENTS.md	General spec
Claude Code	CLAUDE.md	Adapted for Claude's tool invocation patterns
Gemini CLI	GEMINI.md	Adapted for Gemini's tools and context window
Kiro	AGENTS.md	Kiro automatically reads AGENTS.md

2. Add Materials

Place materials you want to digest into the appropriate raw/ subdirectory:

# Web articles (recommended: save as Markdown using Obsidian Web Clipper)
raw/articles/2026-04-13-some-article.md

# Academic papers
raw/papers/2026-04-13-some-paper.pdf

# Meeting notes, reading notes
raw/transcripts/2026-04-13-meeting-notes.md

It's recommended to note the source URL, author, and date at the beginning of each file.

3. Let the LLM Digest

Tell your LLM Agent:

Digest raw/articles/2026-04-13-some-article.md

The LLM will:

1. Read the material and discuss key points with you
2. Create a material summary page (wiki/summaries/)
3. Create or update related entity and concept pages
4. Maintain all cross-references
5. Update the global index and operation log

4. Ask Questions

Ask the LLM directly — it will answer based on compiled knowledge in the wiki:

Compare the pros and cons of Playwright vs Cypress

Valuable answers can be archived as wiki pages (comparisons or synthesis).

5. Publish Deliverables

Produce audience-facing deliverables from wiki content:

Turn the wiki knowledge about Playwright vs Cypress into a blog post

The LLM will distill content from the wiki, convert it to standalone standard Markdown, and save it to output/posts/. You review and finalize.

Supported output types: blog posts, research reports, Marp slides, tutorials, knowledge briefings.

6. Health Check

Periodically have the LLM check wiki health:

Run a health check on the wiki

The LLM will check for contradictions, orphan pages, missing concept pages, outdated information, and more.

Obsidian Configuration Guide

This project's wiki layer is fully compatible with Obsidian. Obsidian is recommended as the browser. The LLM edits on one side while you browse in real-time in Obsidian.

Basic Configuration

1. Open this project's root directory as a Vault in Obsidian

2. Go to Settings → Files and links and adjust the following:

   Setting	Recommended Value	Description
   Default location for new notes	raw/articles	New notes default to the raw directory
   New link format	Shortest path when possible	Concise link format
   Use [[Wikilinks]]	Enabled	Enable wikilink syntax
   Attachment folder path	raw/assets/	Centralized attachment storage
   Detect all file extensions	Enabled	Recognize non-md files like PDFs

3. Go to Settings → Editor:

   Setting	Recommended Value	Description
   Show frontmatter	Enabled	Display YAML metadata
   Readable line length	Enabled	Improve reading experience

Recommended Plugins

Obsidian Community Plugins

Search and install in Settings → Community plugins:

Plugin	Purpose	Configuration Notes
Dataview	Dynamic querying of page metadata	Use dataview code blocks to query frontmatter fields like tags, sources, etc., generating dynamic tables
Excalidraw	Hand-drawn style whiteboard and diagrams	Embed architecture diagrams, flowcharts, concept relationship maps in wiki pages. Store files in raw/assets/, embed via ![[filename.excalidraw]]
Local Images Plus	Auto-download remote images to local	Set download path to raw/assets/. When pasting or importing articles with remote images, automatically downloads images locally and replaces links, preventing broken external links

Browser Extensions

Extension	Browser	Purpose	Configuration Notes
Obsidian Web Clipper	Chrome / Firefox / Safari	One-click save web articles as Markdown	After installation, set save path to raw/articles/. Works with Local Images Plus to auto-download images from articles

Graph View Configuration

Obsidian's Graph View is the best way to browse knowledge structure:

1. Open Graph View (Ctrl/Cmd + G)
2. Recommended filter settings:
   • Filters → Search files: Enter path:wiki/ to show only wiki pages
   • Groups: Color-code by directory to distinguish entities, concepts, summaries, etc.
     • path:wiki/entities → Blue
     • path:wiki/concepts → Green
     • path:wiki/summaries → Yellow
     • path:wiki/comparisons → Orange
     • path:wiki/synthesis → Purple
3. Observe:
   • Nodes with the most connections are knowledge hubs
   • Isolated nodes may need additional cross-references
   • Clusters reveal natural groupings of knowledge domains

Dataview Query Examples

After installing the Dataview plugin, you can use the following queries in any Markdown file:

List all pages by tag:

```dataview
TABLE tags, description, updated
FROM "wiki"
WHERE contains(tags, "testing")
SORT updated DESC
```

List recently updated pages:

```dataview
TABLE category, description, updated
FROM "wiki"
SORT updated DESC
LIMIT 10
```

List all material summaries with their sources:

```dataview
TABLE sources, description, created
FROM "wiki/summaries"
SORT created DESC
```

Downloading Images Locally

To prevent external image links from breaking, it's recommended to download article images locally:

• Recommended: Install the Local Images Plus plugin (see recommended plugins above). It automatically downloads remote images to raw/assets/ and replaces links when you paste or import articles — fully automatic.
• Manual: Go to Settings → Hotkeys, search for "Download", and bind a shortcut (e.g., Ctrl+Shift+D) for "Download attachments for current file". After saving an article with Web Clipper, press the shortcut.

License

This project's structure and schema design is based on Karpathy's LLM Wiki concept

MIT License — see LICENSE.