The knowledge graph

Every link between documents is an edge. Memory and knowledge here are a graph the agent walks, not a pile of files it greps.

Knowledge graph of soul.demarkus.io: 261 documents and 840 edges, clustered by prefix, with typed relations highlighted

This project's memory, crawled and drawn by an agent in one mark_graph call

Links are edges

Write a markdown link and you have an edge. Each edge keeps where it came from: the label, the section it sits in, how often it appears. Every publish updates the graph.

A new document is linked from its hub on first publish, so nothing starts orphaned. The doctor reports broken links, orphans, and missing hubs.

Typed relations

Say why two documents connect with rel-<predicate> metadata:

rel-supersedes: /adr/0002-metadata-alignment.md
rel-depends-on: /architecture.md, mark://other-world/spec.md

Supersedes, depends on, implements, derives from: whatever your catalog needs. A typed relation and a plain link between the same pair are separate edges.

The agent walks it

Tool	What it does
`mark_lookup`	Find by tag, title, or section text. With a `budget`, returns the matching passages plus the documents related to the best match
`mark_explore`	Orient around one document: outline, links out, backlinks, siblings, grouped by relation
`mark_backlinks`	Who links here, revision-aware, with freshness checks
`mark_graph`	Crawl outbound links to a depth; typed relations included

The graph persists between sessions. What the agent learned yesterday about how documents connect is still there today.

Across worlds

In a knowledge system the server derives the cross-world graph itself and publishes it to the hub (/graph.md, /graph/worlds/). A link from one team’s world into another’s is an edge like any other. No crawler to deploy.

See it

Library: the floor shows every world; each world has a map and a link-graph canvas.
TUI: in demarkus-tui, d shows links out, r backlinks and typed relations, t the whole crawled graph. Keys →
Export: mark_graph_export writes the graph as markdown with mark:// links; mark_graph_publish publishes it for other agents to crawl.

Why a graph

Less context. The agent follows the edge it needs instead of loading every file.
Better answers. Related decisions surface with the match, not after the mistake.
Honest history. Superseded decisions stay linked to what replaced them.