Agent Wiki Idea

Exploring what a wiki looks like when its primary readers are AI agents — vector-native, tool-driven, no pages.

About This Project

Agent-Wiki-Idea is a planning and notes repository that investigates what a knowledge base would look like if its median reader were an LLM agent instead of a person. It contrasts this premise with traditional wikis such as MediaWiki, DokuWiki, Notion, and Confluence, which are built around hierarchical navigation, page titles, and keyword search. The project proposes that the atomic unit of knowledge be a typed chunk (such as a fact, procedure, or decision) stored with embeddings, with pages treated as on-demand views rather than authored artifacts.

The working hypotheses in the repository cover several design areas: writes performed through capability tools such as add_fact, add_procedure, update_status, and deprecate; reads limited to a semantic search(query, filters) interface returning ranked, cited chunks; mandatory provenance metadata including author, timestamp, source, confidence, and supersession links; and freshness enforcement through TTLs or review dates. The document also lists open questions for further work, including schema requirements, conflict resolution between contradictory writes, signed authorship, semantic deduplication, governance of authoritative content, migration of existing human wikis, and evaluation methods when no human reads the output. The repository currently contains only documentation, with a notes/ directory serving as the scratchpad, and no code.

The project is relevant to developers building agent runtimes, retrieval-augmented generation systems, or MCP-based tooling who need to decide how knowledge should be stored, written, and trusted by automated consumers. It may also interest researchers examining provenance, trust, and freshness in machine-consumed documentation.

Project Details

Explore More Projects