Skip to content

Repository files navigation

LLM Wiki — a second brain for AI agents

Turn PDFs, articles, transcripts, and notes into a shared wiki that your AI agents can search, cite, and keep up to date. Add a source once. Ask questions later. Keep the useful answers.

Works with Claude Code, Codex, Cursor, Gemini CLI, OpenCode, OpenClaw, Pi, and OMP. Read the documentation.

What is LLM Wiki?

AI agents are good at the task in front of them, but a new session starts with limited context. LLM Wiki gives them a shared memory that lives inside your project.

When you add a source, the agent turns it into linked Markdown pages. Later, it can find the right section and answer with citations. Useful answers can be saved back into the wiki, so the knowledge grows instead of being rebuilt from scratch.

Everything canonical stays in readable Markdown. Default semantic search is local—no hosted vector database or embedding service.

What's new in v3.0.0

  • Local semantic search by default. FastEmbed runs BAAI/bge-small-en-v1.5 on-device; sqlite-vec stores derived vectors in each wiki's cache.
  • Hybrid precision and recall. BM25 exact-term ranks and semantic ranks are fused with RRF.
  • Incremental indexing. Content hashes limit re-embedding to new or changed sections and remove deleted sections.
  • Zero provider surface. No OpenAI-compatible endpoint, API key, remote text transfer, provider consent, or per-query charge.
  • Safe lexical escape hatch. --no-embed stays pure-Python BM25; local backend failures fall back to it automatically.
  • No content migration. Existing Markdown remains valid and old embeddings.jsonl caches are simply ignored.

Why use it?

  • Stop repeating project context. Your agent can read the knowledge you already collected.
  • Use the same wiki across agents. The files are plain Markdown, not tied to one model or tool.
  • Trace every answer. Citations point back to the wiki page and original source.
  • Keep the wiki healthy. The agent updates links, summaries, and contradictions as new sources arrive.
  • Stay in control. Your pages remain readable, editable, and versionable.

LLM Wiki works best for knowledge that grows over time: research, meeting notes, customer calls, papers, articles, and project decisions. Use a regular database when your main problem is structured records and transactions.

Installation

uv is the only prerequisite: it creates the pinned script environments on every supported platform. Install it using the official uv instructions before running /wiki:init or the natural-language equivalent.

Plugin/skill installation copies all agent-facing commands and bundled tools. Wiki initialization then performs mandatory runtime setup: FastEmbed 0.8.0, sqlite-vec 0.1.9, PyYAML 6.0.3, the local BAAI/bge-small-en-v1.5 model, the parse cache, and vectors for every existing section. Upgrade runs the same setup again, incrementally synchronizing changed and deleted sections.

Claude Code — full plugin

The native path: the skill, the seven /wiki:* slash commands, and the marketplace manifest all ship in one install.

/plugin marketplace add praneybehl/llm-wiki-plugin
/plugin install llm-wiki@llm-wiki

Once installed, the plugin works in any project — the wiki itself lives in the project's working directory, not in the plugin.

Other coding agents — skill only

The llm-wiki skill uses the standard agentskills.io format, so it installs cleanly into any agent supported by the skills CLI. Pick the --agent flag that matches your setup:

# Install globally so the skill is available across all projects
npx skills add praneybehl/llm-wiki-plugin -a <agent> -g

# Or install into the current project only
npx skills add praneybehl/llm-wiki-plugin -a <agent>
Agent --agent value Invoke via Scripts run
Claude Code claude-code /wiki:* slash commands (bundled) or natural language
Codex (OpenAI) codex /skills or $llm-wiki / natural language
Cursor cursor /llm-wiki or natural language
Gemini CLI gemini-cli /skills management commands / natural language
OpenCode opencode natural language (agent invokes the native skill tool)
OpenClaw openclaw auto-exposed as a user command
Pi Agent pi /skill:llm-wiki or natural language
OMP ("Oh My Pi") manual (see below) natural language (skills auto-surface via skill://)

OpenCode also reads .claude/skills/ and ~/.claude/skills/, so if you already installed the skill for Claude Code you can use it in OpenCode without a second install.

Hermes Agent (Nous Research), OMP ("Oh My Pi"), and other agentskills.io-compatible runtimes that aren't yet in the npx skills registry can still use this skill — clone the repo and symlink or copy skills/llm-wiki/ into the agent's skills directory. Hermes reads from ~/.hermes/skills/; OMP reads managed/user skills from ~/.omp/agent/skills/ and surfaces them via skill://.

git clone https://github.com/praneybehl/llm-wiki-plugin.git
mkdir -p ~/.hermes/skills ~/.omp/agent/skills
ln -s "$(pwd)/llm-wiki-plugin/skills/llm-wiki" ~/.hermes/skills/llm-wiki
ln -s "$(pwd)/llm-wiki-plugin/skills/llm-wiki" ~/.omp/agent/skills/llm-wiki

A few things to know when using the skill outside Claude Code:

  • Slash commands are Claude Code-only. The seven /wiki:* commands live in commands/wiki/ as Claude Code plugin manifests. In other agents, invoke the skill by natural language ("add this paper to the wiki", "what does the wiki say about X", "lint the wiki") — the SKILL.md handles the rest.
  • All bundled tools are agent-accessible. Every listed agent can invoke setup_wiki.py, hybrid/lexical search, lint, stats, graph lint/extract/query, and initialization through the installed skill. Dependency-bearing scripts carry pinned PEP 723 metadata and run with uv run --script; initialization and upgrade verify the full runtime before reporting readiness.
  • The wiki itself is agent-agnostic. It's just a directory of markdown files. You can ingest with one agent and query with another; nothing in wiki/ ties it to a specific runtime.

Quick start

In a project where you want to keep a wiki:

/wiki:init

Already have a wiki from an earlier plugin version? Run /wiki:upgrade instead. Init and upgrade both install and verify the complete pinned local runtime, cache the embedding model, build the parse cache, and synchronize all current vectors. They add missing wiki files idempotently and never overwrite existing pages or SCHEMA.md.

Initialization bootstraps wiki/ and raw/ with SCHEMA.md, index.md, log.md, and a page template, then emits a JSON runtime report. Setup is complete only when it reports "status": "ready".

As part of the same step, the skill will propose wiring the wiki into your project's agent-memory file so the agent remembers it in future sessions without being told. The target file depends on which agent you use: CLAUDE.md for Claude Code, AGENTS.md for Codex / Cursor / OpenCode / Pi / OpenClaw / OMP, GEMINI.md for Gemini CLI. If you run multiple agents in the same project, use AGENTS.md and symlink CLAUDE.md to it. The skill never writes to a memory file without your approval — see skills/llm-wiki/references/agent-memory-integration.md for the canonical stanza and a three-line short variant.

Drop your first source into raw/ (a PDF, a markdown clipping from the Obsidian Web Clipper, a transcript, anything textual), then:

/wiki:ingest raw/your-source.pdf

Claude reads the source (chunk-reading if it's large), discusses the takeaways with you, writes a summary page, identifies which existing entity and concept pages it touches, surgically updates them with str_replace, creates new pages for new entities and concepts (each with at least one inbound link), updates the index, and appends to the log.

Once you have a few sources in:

/wiki:query What does my wiki say about X?

Claude reads the index to find candidate pages, reads them, synthesizes an answer with [[wikilink]] citations, and offers to file substantive answers back into wiki/synthesis/ so the exploration compounds.

Periodically, run a health check:

/wiki:lint

This catches orphan pages, broken wikilinks, oversized pages, missing frontmatter, and stale claims. Findings are presented as proposed edits — Claude doesn't rewrite your wiki silently.

Commands

Command What it does
/wiki:init Bootstrap a new wiki structure in the current project (now includes wiki/graph/).
/wiki:ingest <source> Process a new source into the wiki; refreshes the graph layer when present.
/wiki:query <question> Answer a question from the wiki with citations; consults graph.sqlite for relational questions when available.
/wiki:lint Structural and semantic health check; also runs graph lint when wiki/graph/ontology.yaml exists.
/wiki:stats Show wiki size, link density, and which scaling threshold the wiki is at.
/wiki:graph <action> extract / lint / neighbors / edges / path / facts against the compiled graph layer.
/wiki:upgrade Upgrade an existing wiki to the current plugin version (idempotent file ops + walked SCHEMA.md merge).

You don't have to use the commands — Claude triggers the underlying skill on natural-language requests too ("add this paper to my wiki", "what does the wiki say about diffusion", "lint the wiki"). The commands are there when you want explicit invocation.

Architecture

The wiki has three layers and three operations.

The three layers are raw sources (your immutable source material in raw/), the wiki (LLM-owned markdown pages in wiki/, organized into sources/, entities/, concepts/, synthesis/), and the schema (wiki/SCHEMA.md, the configuration file that documents the conventions for this specific wiki and that you co-evolve with Claude over time).

The three operations are ingest (compile a new source into the wiki, touching ten or fifteen pages in one pass), query (navigate the wiki to answer a question, with the option to file the answer back as a new synthesis page), and lint (periodic health check on structural and semantic integrity).

There is also an optional fourth layer: the graph at wiki/graph/. Pages can declare typed graph: metadata in frontmatter (e.g. founded, proposed, depends_on) with an explicit source-page slug and evidence quote. A bundled extractor compiles every page into nodes.jsonl, edges.jsonl, graph.sqlite, and graph.graphml. Markdown stays canonical — the graph can be deleted and rebuilt at any time. Use /wiki:graph for relational queries that the index alone can't answer cheaply ("what's connected to X", "who proposed Y", "shortest path A → B").

For the full architecture write-up including page types, frontmatter conventions, the rationale behind each design choice, and the workflow procedures, see the reference docs inside the skill at skills/llm-wiki/references/. The graph layer is documented in skills/llm-wiki/references/graph-workflow.md.

How it scales

The single biggest failure mode of the LLM Wiki pattern is the wiki becoming a context bottleneck as it grows. The plugin's design is shaped largely around avoiding that failure.

Atomic pages. Every page is about one concept. Soft cap 400 lines, hard cap 800 lines. The lint script enforces this. Any single page read is bounded, so even at a thousand pages a query that reads five pages stays well within context.

Sharded indexes. When wiki/index.md exceeds 300 lines or the wiki passes ~150 pages, the index shards into per-type files under wiki/indexes/. The top-level index becomes a tiny directory of shards. Reading the index is now a two-step lookup but each step is bounded.

YAML frontmatter on every page. The wiki_search.py script can filter on type, tags, and updated without reading page bodies, so retrieval scales independently of total wiki size.

Index-first navigation. Queries read the index first, identify candidate pages from one-line summaries, and only read those. The index is the cache that makes the whole pattern scalable.

Local hybrid search. When index summaries are insufficient, uv run --script skills/llm-wiki/scripts/wiki_search.py "query terms" ranks Markdown sections with local FastEmbed semantic vectors plus BM25, fused through RRF. sqlite-vec stores content-hashed vectors in wiki/.wiki-cache/embeddings.sqlite; only changed sections are re-embedded. No API key, provider, or outbound wiki/query text is involved. --json emits evidence rows and --cache persists parsed sections. For the genuinely dependency-free path, invoke python skills/llm-wiki/scripts/wiki_search.py "query terms" --no-embed; --granularity page restores whole-page ranking.

Surgical edits. When updating a page during ingest, Claude uses str_replace to touch the relevant section, not rewrite the page. Fast, token-efficient, and preserves diff quality if you keep the wiki in git.

Chunked source ingestion. Large raw sources (long PDFs, multi-hour transcripts) are read in chunks during ingest, not loaded whole. The ingest workflow handles this automatically.

The /wiki:stats command surfaces which threshold you're at and what migration to do next. The full scaling playbook with thresholds and migration procedures is at skills/llm-wiki/references/scaling-playbook.md.

What it doesn't try to be

This is a markdown-based personal knowledge base, not a database. If your domain is fundamentally relational — tracking 100,000 customer records, complex org structures, financial ledgers with strict referential integrity — a real database serves better. Markdown plus grep plus frontmatter scales further than people expect (Karpathy's own wiki was about 100 articles and 400K words working fine), but at some point the right tool changes.

Failure modes to know about, all discussed in skills/llm-wiki/SKILL.md: silent corruption (a misreading of one source becomes an authoritative-looking wiki page that influences subsequent ingests — mitigated by mandatory sources: frontmatter and lint), the wiki-reads-its-own-output drift (Claude treating prior wiki pages as ground truth rather than checking against raw sources — mitigated by the ingest workflow's instruction to re-read raw sources when updating claims), and the maintenance ratchet (lint reports growing faster than you can review — mitigated by the scalability discipline, but if it happens, the schema needs revision).

Integrations

The skill is the agent-side install — every adapter table above gets you that. For specific platforms there's also an optional human-side companion that brings the wiki into view where operators already work.

Paperclip — paperclip-plugin-llm-wiki

For teams that run their work in Paperclip, an optional plugin at integrations/paperclip/plugin/ surfaces the wiki inside Paperclip's UI. Five surfaces, all read-only:

Surface What it does
Wiki sidebar Browse the wiki by type, drill into pages, search across the whole wiki
Full-page view The same browser at full width — for reading multiple linked pages
Issue context tab Top wiki pages relevant to the open issue, ranked by BM25 over title + description
Dashboard health widget Page count, lint status (pass / warn / fail), link density, sharding-threshold messages
wiki.query agent tool BM25 search via tool call — for HTTP-only adapters that don't run the skill directly

Install (once v0.1 ships to npm):

pnpm paperclipai plugin install paperclip-plugin-llm-wiki

The plugin is read-only by design and pairs with the skill. The skill writes the wiki on heartbeat; the plugin reads it from inside Paperclip. Editing still happens through agents on heartbeat or the operator's editor of choice (Obsidian, Claude Code, direct SSH).

Documentation:

Tooling

Once the wiki is set up, you can read and edit pages with any markdown viewer. Obsidian is a particularly good fit because of its graph view, [[wikilinks]] syntax, and Web Clipper extension, but it isn't required — the wiki is just a directory of markdown files in your project.

The tools require Python 3.10+ and uv. Initialization and upgrade invoke setup_wiki.py, whose PEP 723 environment pins FastEmbed 0.8.0, sqlite-vec 0.1.9, and PyYAML 6.0.3; it caches the model and synchronizes the complete wiki index. Search, graph lint, and graph extraction also carry their own pinned script metadata. Lexical wiki_search.py --no-embed, wiki_lint.py, wiki_stats.py, and graph queries remain directly runnable with stdlib Python. All eight scripts live in skills/llm-wiki/scripts/ after install.

Credits

The pattern is from Andrej Karpathy's llm-wiki gist, April 2026. This plugin is one implementation of the pattern; others worth looking at include vanillaflava/llm-wiki-claude-skills, skyllwt/OmegaWiki, and axoviq-ai/synthadoc.

The v2 retrieval architecture was informed by Cerebras's engineering article, How We Built Our Knowledge Base, especially its section-level retrieval, multi-retriever fusion, evidence packaging, and evaluation approach.

The v3 local retrieval cutover was evaluated against Basic Memory. LLM Wiki adopts the same local-first shape—FastEmbed, SQLite vector search, and incremental derived indexes—while keeping its existing Markdown schema, typed graph, RRF ranking, and agent-native workflows instead of adding Basic Memory's application and MCP service as a second canonical system.

License

MIT — see LICENSE.

Contributing

Issues and PRs welcome. The skill itself is at skills/llm-wiki/; the slash commands at commands/wiki/. Before opening a PR, please run the bundled scripts against a small test wiki to verify nothing broke.

Releases

Packages

Contributors

Languages