LLM Wiki

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.