AK LLM Wiki
A plugin for building and maintaining a persistent, compounding knowledge base
as interlinked markdown files. Based on Andrej Karpathy's LLM Wiki pattern.
Unlike RAG (which rediscovers knowledge from scratch per query), the wiki is
compiled once and kept current. Cross-references are already there.
Contradictions have already been flagged. You curate sources and ask good
questions; the agent does all the bookkeeping — reading, summarizing,
cross-referencing, filing, maintaining consistency. You (almost) never write
wiki pages yourself.
Obsidian is the IDE. The LLM is the programmer. The wiki is the codebase.
— Karpathy
Design lineage
This plugin implements Karpathy's original concept and extends it with three
systems he left open. The table below shows which ideas come from the original
and which are plugin extensions.
From Karpathy's original
| Idea | Original quote / reference |
|---|
| Compiled once, kept current (not RAG) | "Unlike RAG... the wiki is compiled once and then kept up to date" |
| Human curates, LLM maintains | "You curate sources and ask good questions, the LLM does the rest" |
raw/ immutable layer | "You don't modify raw source material" |
index.md — read index first | "The LLM reads the index first to find relevant pages" |
log.md — append-only action log | "A chronological log of all actions taken" |
| Ingest: 10–15 pages per source | "A single source can touch 10-15 pages" |
| Query results file back and compound | "Queries compound in the wiki just like ingested sources" |
| Lint: stale content + data gaps | "Lint also finds data gaps that could be filled with a web search" |
| Schema co-evolves with the wiki | "The schema... co-evolves along with the wiki over time" |
| Git repo by default | "The wiki directory is just a git repo" |
| Obsidian as the IDE | "Obsidian is the IDE. The LLM is the programmer." |
| Everything optional and modular | "Pick what's useful, ignore what isn't" |
| qmd for scaling search | "qmd... for search at scale" |
| Web search to fill gaps | "The LLM is good at suggesting new questions to investigate" |
Plugin extensions (this project)
| Extension | What it adds | Why |
|---|
| SCHEMA.md as authoritative config | All conventions in one file; agent reads and enforces rather than hardcoding | Karpathy said schema co-evolves — we formalized where it lives and who enforces it |
| Interactive domain-specific init | wiki-init proposes categories per domain (research/book/business/personal) | Karpathy's structure is intentionally abstract — a real tool needs a concrete bootstrapper |
Bulk import (wiki-import) | 5-phase migration from Obsidian/Notion/Confluence/folders with checkpoint/resume | Original assumes you start fresh; most users have existing notes |
Image handling in wiki-ingest | Auto-download referenced images to raw/assets/, rewrite to local paths | URLs rot; Karpathy didn't specify image handling |
Structured graph queries (wiki-graph) | Frontmatter filters, BFS neighbors, shortest-path, hubs/orphans — no persistent graph DB | The [[wikilink]] graph is implicit in Karpathy's design; we made it queryable |
| Custom frontmatter dimensions | Domain-specific fields (e.g. version:) declared in SCHEMA.md, prompted during ingest | Karpathy mentions frontmatter but doesn't specify extensibility |
| Three-tier memory aging | active/archived/frozen tiers computed on-the-fly from source dates | Karpathy notes "stale content" in lint but doesn't propose a temporal model |
| External fallback plugins | .wiki-plugins.yaml registers CLI tools (deepwiki-cli, web search) as wiki-query fallback | Karpathy mentions web search for gaps; we generalized to any external tool with a closed-loop ingest pipeline |
| Multi-format query output | markdown / table / slides (Marp) / chart (matplotlib) / canvas (Obsidian) | Original doesn't specify output formats |
| Plugin packaging | .claude-plugin/ + .codex-plugin/ + skill files for Claude Code and Codex CLI | Karpathy's concept is tool-agnostic; this wraps it for specific platforms |
What we deliberately did NOT add
- Persistent graph database — the filesystem is the graph; scanning hundreds of pages takes milliseconds
- Auto-pruning of frozen content — planned as "auto-dreaming" in v2; frozen = parked, not deleted
- Embedding-based semantic search — deferred to qmd; the built-in 3-pass scan handles Karpathy's sweet spot (~100 sources)
- Multi-user access control — the wiki is a git repo; use branches and PRs for collaboration
Installation
Claude Code
# Install via marketplace
claude plugin install litianyi/AK-llm-wiki
surebeli
surebeli
