A Minimalist Knowledge Base for AI Agents

A new open-source project called WUPHF provides a unique approach to knowledge management for AI agents, using markdown files stored in a git repository as its primary data structure. This design prioritizes durability, provenance, and the ability for agents to both read from and write into a shared knowledge base.

The system consists of:

• Private agent notebooks: Each agent gets its own Markdown file where it can draft ideas and notes (e.g., agents/my-agent/notebook/.md)
• Team wiki: A central repository for canonical knowledge accessible to all agents (at team/.)
• Fact logs: Append-only JSONL files that track specific facts associated with entities (e.g., team/entities/company-acme.facts.jsonl)

The architecture includes a draft-to-wiki promotion flow where agent notes can be reviewed and elevated to official wiki content, creating backlinks in the process.

Technical Details

Instead of relying on vector databases or graph solutions, WUPHF uses:

• Markdown for its human-readable and durable storage format
• Bleve (BM25) search engine for efficient keyword lookups
• SQLite to manage structured metadata like entity relationships and access controls

The current benchmark achieves 85% recall@20 on BM25, with SQLite-vector as a fallback if performance degrades.

Key Features

• Git-based versioning ensures all changes are tracked with clear provenance
• Canonical IDs provide stable references to entities across updates
• Automatic linting detects contradictions and broken links
• Slash commands and integrated tools enable seamless information retrieval within agent workflows

The project emphasizes a minimalist approach where the knowledge base evolves organically through agent contributions rather than top-down design.

WUPHF is available on GitHub under an MIT license and can be self-hosted with existing AI agent setups. A demo video showcases how to create, synthesize, and commit knowledge entries using terminal commands.