git:20261029.d74a750 to git:20261029.e5f1a50

# LLM Wiki

> LLM Wiki is an open-source skill and plugin for building an agent-maintained, continuously compiled knowledge base from project sources. It implements Andrej Karpathy's LLM Wiki pattern with canonical Markdown pages, section-level retrieval, optional typed graphs, and support for multiple coding agents.

The canonical project is https://github.com/praneybehl/llm-wiki-plugin. The documentation below is public, pre-rendered as static HTML, and maintained with the current release.

## Start here

- [Overview](https://praneybehl.github.io/llm-wiki-plugin/): What LLM Wiki is, how it differs from query-time RAG, and the three-step quick start.

- [Getting started](https://praneybehl.github.io/llm-wiki-plugin/getting-started.html): Install `uv` and the skill/plugin for Claude Code, Codex, Cursor, Gemini CLI, OpenCode, OpenClaw, Pi, OMP, or Hermes; initialization then installs the complete pinned runtime, local model, and indexes.

- [Commands](https://praneybehl.github.io/llm-wiki-plugin/commands.html): Reference for `/wiki:init` (including `--hosted <url>`), `/wiki:ingest`, `/wiki:query`, `/wiki:lint`, `/wiki:stats`, `/wiki:graph`, `/wiki:migrate`, `/wiki:upgrade`, `/wiki:learn`, and `/wiki:evolve`.

- [Workflows](https://praneybehl.github.io/llm-wiki-plugin/workflows.html): Step-by-step ingest, query, and maintenance behavior.

- [Learning and skill evolution](https://praneybehl.github.io/llm-wiki-plugin/evolution.html): Bounded training, persistent evidence consolidation, whole-skill proposals, validation selection, independent final tests and cross-agent transfer. Includes nine host adapters, durable interrupted traces, calibrated judging, ingestion/graph/hybrid-search checks, paired task-level analysis, pre-request API spending reservations and recovery.

- [WikiSkill research basis](https://arxiv.org/html/2608.27454v1): The source paper for the optional evolution workflow; the evolution guide maps adopted ideas, project-specific choices and measured limitations.

## Retrieval and structure

- [Search and retrieval](https://praneybehl.github.io/llm-wiki-plugin/search.html): Mandatory init/upgrade setup for local FastEmbed + sqlite-vec semantic search fused with section-level BM25, structured JSON evidence, and incrementally synchronized caches.

- [Graph layer](https://praneybehl.github.io/llm-wiki-plugin/graph.html): Optional typed, provenance-backed relationships compiled from canonical Markdown.

## Hosted wiki

- [Hosted wiki](https://praneybehl.github.io/llm-wiki-plugin/hosted.html): Optional one-click Cloudflare deployment: one Worker with a SQLite Durable Object for each wiki, MCP and JSON API access, per-agent tokens or OAuth, atomic commits with history and conflict detection, hybrid search with Workers AI and Vectorize, a browser reader, migration with `wiki_remote.py import`, a pull mirror, Free and Paid plan limits, costs and security.

- [Decision models](https://praneybehl.github.io/llm-wiki-plugin/decisions.html): Optional, off-by-default small decision models (Jev, Liquid d1) for a search relevance gate, graph evidence checks, contradiction screening, ingest shortlists and untrusted-text flags; backends, privacy, licence and fail-open rules.

## Ecosystem

- [Supported agents](https://praneybehl.github.io/llm-wiki-plugin/agents.html): Verified agent support, installation methods, and memory-file conventions.

- [Upgrade to v4](https://praneybehl.github.io/llm-wiki-plugin/upgrade.html): v4 adds the optional hosted wiki and decision layer; local wikis need no migration. Also covers upgrades from v3.1 and earlier.

## Source and licensing

- [GitHub repository](https://github.com/praneybehl/llm-wiki-plugin): Source code, skill files, commands, the hosted server, tests, and release history.

- [License](https://github.com/praneybehl/llm-wiki-plugin/blob/main/LICENSE): MIT.

- [Karpathy's LLM Wiki gist](https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f): The original pattern that inspired this implementation.

- [Cerebras: How We Built Our Knowledge Base](https://www.cerebras.ai/blog/how-we-built-our-knowledge-base): Engineering reference that informed v2 section retrieval, fusion, evidence packaging, and evaluation.