wikiwiki is a project to build tooling and conventions for an LLM-maintained personal knowledge wiki: a directory of interlinked Markdown files that models write and update for you, sitting between immutable raw sources (clips, papers, repos, images) and how you explore and question that material.
The goal is compounding knowledge—ingests, answers, and “lint” passes improve the same persistent wiki over time—instead of one-off chat sessions where nothing durable accumulates.
- Ingest: Drop sources into a
raw/area; agents (or scripts) summarize, extract entities and concepts, cross-link pages, and keep indexes and logs consistent with a project schema. - Query: Ask questions against the wiki; strong answers can be filed back as new pages so your research trail becomes part of the corpus.
- Lint: Periodic health checks—inconsistencies, stale claims, orphan pages, missing links, gaps that web search or new sources could fill.
- Viewing: Markdown-first, local files you own; Obsidian (or similar) as the human-facing “IDE” for graph and preview, matching the pattern described in the references below.
- Optional extras: Small search or CLI tools over the wiki for agents, Marp or other renderers for slides, and room to evolve toward synthetic data or fine-tuning on wiki text later.
Exact repo layout (raw/, wiki root, config/schema filenames) will firm up as implementation lands; the behavioral contract is: you curate sources and questions; the LLM does the grunt work of maintaining the wiki.
| Doc | Role |
|---|---|
| docs/wikiwiki.md | Canonical write-up of the pattern (original post). |
| docs/llm-wiki-pattern-video-wisdom.md | Build-along notes, transcript pointers, and structured takeaways (walkthrough zVEb19AwkqM). |
| docs/i-built-self-evolving-claude-code-memory-with-karpathys-llm-knowledge-bases-video-wisdom.md | Fabric extract_wisdom on I Built Self-Evolving Claude Code Memory w/ Karpathy's LLM Knowledge Bases (7huCP6RkcY4). |
| docs/transcripts/ | Full captions + cleaned transcripts for linked videos. |
Early stage: documentation and reference material first; application code and automation will follow as the design stabilizes.