跪拜 Guibai
← All articles
Backend

A Self-Growing Personal Knowledge Base with LLM Wiki, Obsidian, and Agent Tooling

By 苍何 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

RAG and file-upload chatbots answer questions without accumulating knowledge. An agent-maintained, file-based wiki that preserves source conflicts and builds a cross-linked concept graph turns personal research into a durable, queryable asset that improves with use instead of resetting after every session.

Summary

Most AI knowledge tools answer questions by reassembling fragments on the fly, leaving no lasting structure. The LLM Wiki approach, inspired by Andrej Karpathy's public methodology, instead tasks an agent with continuous, incremental maintenance: new material is compared against existing pages, concepts are extracted into their own notes, and conflicting viewpoints are preserved with source and timestamp rather than overwritten. The system splits into a Raw layer for original sources, a Wiki layer for structured understanding, and a Schema layer (`AGENTS.md`) that governs how the agent archives, links, and handles disputes.

Obsidian serves as the local, plain-text storage and visualization layer, while WorkBuddy or Codex acts as the agent execution layer that reads directories, creates pages, and updates bi-directional links. Three build methods are detailed: a direct prompt-driven setup, an open-source `claude-obsidian` plugin that wraps common operations into agent skills, and the WeSight plugin that embeds the entire workflow directly inside Obsidian's interface. The result is a knowledge base that grows with each ingested article or conversation, where the agent handles the repetitive structuring and the human supplies high-value material and editorial judgment.

Takeaways
Standard RAG and notebook tools reassemble answers from fragments each time, leaving no persistent knowledge structure behind.
An LLM Wiki uses a three-layer architecture: Raw (original sources, read-only), Wiki (AI-maintained structured pages), and Schema (rules in AGENTS.md for archiving, linking, and conflict handling).
When new material enters Raw, the agent searches existing Wiki pages, supplements or creates concept/entity/topic pages, and preserves conflicting viewpoints with source and timestamp rather than overwriting.
Obsidian's local Markdown vault gives the agent direct file-system access for reading, writing, and linking, while keeping data sovereign and version-controllable via Git.
Three build methods range from a raw prompt-driven setup to the claude-obsidian open-source plugin to the WeSight plugin, which integrates ingestion, updating, and retrieval entirely inside Obsidian's UI.
The agent handles repetitive structuring and maintenance; the human supplies high-value material, sets rules, and verifies critical conclusions.
Mobile and browser clipping tools (Obsidian Web Clipper, WorkBuddy Mini Program) feed Raw from anywhere, closing the capture-to-structure loop.
Conclusions

LLM Wiki inverts the dominant RAG paradigm: instead of retrieving fragments to assemble ephemeral answers, it builds a persistent, incrementally updated knowledge graph that becomes a reasoning substrate for future queries.

Obsidian's plain-text, file-system-native design is not just a preference for local-first purists; it is an architectural enabler that lets an agent operate on knowledge as files, with full Git history and no API lock-in.

The explicit handling of conflicting information—preserving source, time, and scope rather than synthesizing a single summary—addresses a failure mode in AI summarization where nuance is quietly erased.

Packaging the workflow into an Obsidian plugin (WeSight) shifts the interaction model from 'developer prompts an agent in a terminal' to 'knowledge work happens inside the note-taking environment,' lowering the barrier for sustained use.

The system's value compounds: each ingested article enriches the concept and entity graph, so later queries resolve faster and with richer context, making the knowledge base genuinely 'self-growing' rather than just searchable.

Concepts & terms
LLM Wiki
A knowledge management architecture where an AI agent continuously maintains a structured wiki from raw source materials, creating concept, entity, and topic pages with bi-directional links, and preserving conflicting viewpoints with provenance rather than synthesizing a single answer.
Raw / Wiki / Schema layers
The three-tier structure of an LLM Wiki: Raw stores original materials read-only; Wiki holds AI-organized structured pages; Schema (typically an AGENTS.md file) defines the rules for how the agent archives, links, cites, and handles conflicts.
AGENTS.md
A Markdown file placed in the root of an Obsidian vault (or project) that serves as a persistent instruction set for AI coding agents, defining rules for file handling, linking, conflict resolution, and maintenance procedures.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗