LLM Wiki: Turning Your AI Agent Into a Knowledge Base That Compounds

What LLM Wiki actually is

LLM Wiki is a tool by nvk for building knowledge bases that an AI agent compiles and maintains over time. It ships in a few forms: a Claude Code plugin, an OpenAI Codex plugin, an OpenCode instruction file, and a portable `AGENTS.md` that drops into any agent able to read and write files and search the web. It is Obsidian-compatible, and everything it produces is plain Markdown you own.

The core idea is that single chat sessions throw away most of what you learn. LLM Wiki tries to fix that. Each run compounds: raw sources get ingested and stay immutable, articles synthesize on top of them with cross-references and confidence scores, and outputs like reports, slide decks, study guides, and implementation plans get filed back into the wiki so the next output builds on the last.

Installation is a one-liner in Claude Code (`claude plugin install wiki@llm-wiki`), a marketplace add in Codex, or an instruction-file path in OpenCode. There is also a Pi setup the project calls out as best for local models, claiming its small system prompt leaves room for the full skill on 32K-context local models.

What it does in practice

The feature list is broad, so it helps to group it. The headline capability is parallel multi-agent research: a single command can dispatch 5 to 10 agents across academic, technical, applied, news, and contrarian angles, with a `--min-time 2h` flag that keeps the run going in rounds and drills into gaps each round surfaces.

There is also a thesis mode that starts from a claim rather than a topic. Agents split across supporting, opposing, mechanistic, meta, and adjacent angles, and the stated output is a verdict rather than a summary, with a second round aimed at fighting confirmation bias. Whether that genuinely reduces bias is something you would want to test on your own material, but the structure is at least more deliberate than a single prompt.

Ingestion covers URLs, files, PDFs, Git doc repos, MediaWiki dumps, message archives, and Wayback CDX snapshots. Beyond research, the tool tracks an inventory of durable things the wiki should remember (items, entities, open questions, watch items, next actions), indexes large external datasets via manifests and query recipes without copying the data, and archives old topics without deleting them.

A few maintenance pieces stand out for anyone who has watched a knowledge base rot. The Librarian scores articles for staleness and quality using a fast metadata pass and a deeper read for flagged items. An Audit step traces outputs back through raw sources, detects drift, and runs fresh research when local evidence is thin. And a Plan command produces wiki-grounded implementation plans that cite articles as evidence, with `--format rfc|adr|spec` options.

Session handling is deliberately restrained. A default-on hook writes redacted events and Markdown digests under `.sessions/`, and the project repeatedly stresses that it captures context without hoarding full transcripts or turning private chats into topic evidence. Feedback curation only promotes durable corrections and approvals, ignoring generic acknowledgements.

Why it matters for builders

If you run an agent like Claude Code or Codex as part of your workflow, the practical problem is continuity: research and decisions evaporate between sessions. LLM Wiki addresses that with structure rather than a vector database alone. Every directory has an `_index.md` so nothing is scanned blindly, sources stay separate from synthesis, and queries come in quick, standard, or deep tiers with a `--resume` option.

Because the output is plain Markdown in a directory you control, it slots into existing setups instead of locking you into a service. That is the appealing part for automation-minded users: you can pair it with Git, Obsidian, or your own pipelines, and trigger runs from whatever orchestrates your agents. It is not a workflow engine like Zapier, Make, or n8n; it is a research-and-memory layer that those tools could call or schedule around.

The local-model angle is worth noting. The bookmark that pointed us here describes a fully local stack: a machine with lots of RAM running local models, a private gateway to reach them from any device, LLM Wiki for the knowledge bases, and a separate retrieval layer on top. That setup is more involved than installing a Claude Code plugin, and it assumes comfort with self-hosting, but it shows the design intent: keep your sources, synthesis, and memory under your own roof rather than inside a hosted chat app.

Where it fits versus the alternatives

Compared with using the ChatGPT or Claude desktop apps directly, the trade is structure for setup effort. The apps are easier to start; LLM Wiki gives you persistent, cross-referenced, auditable artifacts that compound. Against a plain note vault, it adds active agents that research, ingest, score, and maintain. Against a managed RAG product, it stays portable and file-based, which means more wiring but no vendor lock-in.

The honest caveat is that capability here depends heavily on the agent and model you run it through, and the broad feature surface means a learning curve. It is open source on GitHub, so the cheapest way to judge it is to install the plugin and run one topic wiki end to end.

If you want help wiring LLM Wiki into your agent stack or a local-model setup, browse the provider directory to find someone who can put it to work.