6. Shared wiki — compounding team knowledge

Status: planned (V2) · Difficulty: intermediate

This page is on the roadmap

The wiki ships and runs daily on the maintainer's own deployment. A user-facing walkthrough is pending.

Scenario (planned)

Every conversation your agents have gets saved as a raw entry. Overnight, a cron runs agentx wiki absorb — an LLM reads the new entries, compares them to existing articles, and writes distilled, cited wiki articles the whole team can read tomorrow. After a week you have projects/, people/, decisions/, patterns/ directories with interlinked [[wikilinks]], each article marked public, shared, or private.

Outline (what this page will teach)

• The three wiki modes: unified (default), flat (Karpathy), graph (knowledge graph)
• Ingesting conversations into raw entries automatically
• agentx wiki absorb — the daily compile step; the prompt template per mode
• Cron pattern: wiki-absorb-midnight with onError: ["notify", "disable"] (see Journey 2)
• agentx wiki query <question> — how agents retrieve before answering
• agentx wiki serve — Wikipedia-style web browser at :4200
• Permissions: when to write public vs shared vs private

Today's nearest equivalents

• Wiki CLI — every command: reference/cli → Wiki
• Concepts page — short wiki primer: concepts
• Source — the absorb prompt template lives at src/wiki/prompts.ts; the SKILL.md that drives absorb is at src/wiki/SKILL.md

Contribute

The ideal V2 page walks through a seven-day loop with screenshots: day 1 entries → day 7 articles with wikilinks.