# LLM Wiki

> LLM-compiled knowledge bases for any AI agent, with parallel multi-agent research, thesis-driven investigation, source ingestion, wiki compilation, and artifact generation.

LLM Wiki is an MIT-licensed, open-source tool that turns any LLM agent into a parallel research engine and an append-only, Markdown-native knowledge base. Created by developer nvk and inspired by Andrej Karpathy's LLM wiki concept, it ships as a Claude Code native plugin, an OpenAI Codex marketplace plugin, an OpenCode instruction file, and a portable `AGENTS.md` for any other LLM agent. All data is stored as plain Markdown files you own, with zero runtime dependencies.

## What It Is

LLM Wiki is a command-driven knowledge management system that orchestrates multiple LLM agents to research topics in parallel, ingest and synthesize sources, and compile structured wiki articles. It sits between a research assistant and a personal knowledge base: agents search the web from different angles (academic, technical, applied, news, contrarian), ingest the best sources into an immutable `raw/` layer, and compile synthesized articles with cross-references and confidence scores into a `wiki/` layer. The result is a portable, Obsidian-compatible Markdown vault you can query, audit, and build deliverables from.

## Research and Investigation Modes

LLM Wiki offers three research intensities and a dedicated thesis mode:

- **Standard** (5 agents): academic, technical, applied, news, and contrarian angles
- **Deep** (`--deep`, 8 agents): adds historical, adjacent-field, and data/stats agents
- **Retardmax** (`--retardmax`, 10 agents): skip planning, cast the widest net, ingest aggressively, compile fast
- **Thesis mode** (`--mode thesis`): starts from a specific claim, splits agents across supporting, opposing, mechanistic, meta/review, and adjacent angles, and delivers a verdict (supported / partially supported / contradicted / insufficient evidence / mixed) rather than a summary. With `--min-time`, round two automatically focuses on the weaker side of the evidence to counter confirmation bias.

The `--plan` flag decomposes a broad topic into 3–5 parallel research paths, confirms with the user, then dispatches all paths simultaneously. The `--min-time` flag keeps research running in rounds—each round drilling into gaps the previous one found—for durations like `1h` or `2h`.

## Architecture and Storage Model

The hub (`~/wiki/`) is a lightweight registry with no content. All content lives in isolated topic sub-wikis under `topics/`. Each topic wiki contains:

- `raw/` — immutable ingested sources (URLs, files, PDFs, Git repos, MediaWiki dumps, Wayback CDX snapshots)
- `wiki/concepts/`, `wiki/topics/`, `wiki/references/` — compiled, synthesized articles
- `inventory/` — durable operational state: items, source candidates, corpora, open questions, next actions
- `inventory/ideas/` — rough proposals being researched and shaped before promotion to Projects
- `datasets/` — manifests for large or external data that stays outside the wiki
- `output/` — generated artifacts (reports, slides, plans, glossaries)
- `schema.md` — human-owned topic guide for local vocabulary and conventions
- `.obsidian/` — Obsidian vault config for graph view and backlinks

Every directory has an `_index.md`; agents read indexes first and never scan blindly. Cross-references use dual-link format (`[[wikilink]]` plus standard Markdown) so the wiki works in Obsidian, Claude Code, GitHub, and plain text editors alike.

## Session Memory and Feedback Curation

LLM Wiki includes a default-on session memory layer under `HUB/.sessions/`. Trusted hooks from Claude Code, Codex, OpenCode, and Gemini-style adapters write redacted checkpoints, compact Markdown digests, per-session state JSON, and feedback candidates—without storing full transcripts. The `/wiki:session rehydrate` command returns a compact context block so a future turn can pick up where the previous one left off. Feedback candidates (corrections, preferences, approvals, plan acceptance) accumulate under `.sessions/feedback/` and are promoted into topic `raw/notes/` only by explicit user action. Users can opt out entirely with `/wiki:session disable`.

## Update: v0.24.4 — Adapter CLI Compatibility

The latest release is **v0.24.4**, published 2026-08-27. It adds `llm-wiki adapter show <id> --json` as an explicit alias for the command's existing JSON output, removing a fresh-session failure mode when an agent follows the adapter workflow using the common `--json` form. The deterministic CLI is synchronized across the Claude, Codex, and OpenCode packages. Notable recent milestones in the changelog include:

- **v0.24.0** — Project Knowledge Checkpoints Export: comprehensive cross-topic handoffs with dry-run-first create, deterministic privacy sealing, and read-only verification
- **v0.23.0** — Personal specialist framework: user-owned, instruction-only `SKILL.md` review methods under `HUB/.skills/` with explicit per-topic allowlists
- **v0.22.0** — Declarative private-adapter routing with provider-neutral intent and exact-URL routes
- **v0.18.0** — Hub-wide portfolio view across active topic wikis

The repository reports 1,127 stars and 108 forks as of the last crawl, per GitHub metadata.

## Features
- Parallel multi-agent research (5, 8, or 10 agents)
- Thesis-driven investigation with verdict output
- Source ingestion (URLs, files, PDFs, Git repos, MediaWiki dumps, Wayback CDX snapshots)
- Wiki compilation with cross-references and confidence scores
- Token-efficient read-only query mode
- Session memory with redacted checkpoints and rehydration
- Feedback curation and explicit promotion
- Project Knowledge Checkpoints Export with privacy sealing
- Personal specialist framework (instruction-only SKILL.md methods)
- Private adapter protocol with scoped resources and plan-bound approval
- Collector catalogs for media, memes, tools, and entities
- Inventory tracking for items, candidates, corpora, and next actions
- Dataset manifests for large external data
- Idea incubation and explicit Project promotion
- Hub-wide portfolio view
- Archive and restore topic wikis
- Artifact generation (reports, slides, study guides, plans, glossaries)
- Obsidian-compatible dual-link format
- Zero runtime dependencies
- Audit and truth-seeking across wiki, outputs, and provenance
- Librarian staleness and quality scoring
- Lint and auto-repair structural issues
- Chrome extension for bounded tab control via MCP tools

## Integrations
Claude Code, OpenAI Codex, OpenCode, Pi / DS4 local models, Obsidian, Chrome (via browser extension), Grok MCP (for tweet ingestion), tobi/qmd (local search engine), iCloud Drive, GitHub, MediaWiki, Wayback Machine CDX API

## Platforms
MACOS, LINUX, API, BROWSER_EXTENSION, VSC_EXTENSION, CLI

## Pricing
Open Source

## Version
v0.24.4

## Links
- Website: https://llm-wiki.net/
- Documentation: https://llm-wiki.net/#usage
- Repository: https://github.com/nvk/llm-wiki
- EveryDev.ai: https://www.everydev.ai/tools/llm-wiki