nvk/llm-wiki

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

What it solves

llm-wiki provides a way for AI agents to build, maintain, and query structured knowledge bases (wikis) from various sources. It solves the problem of agents having limited memory or context windows by allowing them to perform deep, autonomous research, ingest sources, and compile a persistent, Obsidian-compatible knowledge base that can be shared across different AI runtimes.

How it works

The project implements a behavioral protocol that can be installed as a plugin for various AI agents (Claude Code, OpenAI Codex, OpenCode, Pi, and others). It allows agents to use a set of commands to perform tasks such as:

  • Research: Launching parallel multi-agent research sessions to investigate topics, decompose questions into playbooks, or conduct thesis-driven investigations (finding evidence for and against a verdict).
  • Ingestion: Importing content from URLs, GitHub repositories, MediaWiki dumps, CSV message archives, and archived snapshots via the Wayback Machine.
  • Curation: Capturing high-signal corrections and preferences as feedback, and using "topic guides" (schema.md) to organize knowledge.
  • Querying: Providing a read-only query protocol for fast lookups and evidence checks within the compiled corpus.
  • Inventory & Datasets: Tracking sources, assets, or external datasets that remain external to the wiki.

Who it’s for

Users who utilize AI coding agents or general-purpose LLM agents and want them to their agents to have a persistent, structured, and verifiable knowledge base that they can autonomously research and research and maintain.

Highlights

  • Multi-Runtime Support: Works across Claude Code, OpenAI Codex, OpenCode, and Pi, with a portable AGENTS.md for any other agent.
  • Curation Tools: Includes a feedback curator for capturing high-signal corrections and automated session capture for context rehydration.
  • Thesis-Driven Research: Specialized modes for investigating specific hypotheses with a verdict verdict.
  • Obsidian-Compatible: The resulting knowledge base is stored in a format compatible with Obsidian.
  • Extensible Ingestion: Supports bulk import of spec repos, MediaWiki dumps, and historical web snapshots.