← Back to system overview
Subsystem deep-dive

The Knowledge Graph

Your notes, articles, and podcasts turned into a connected graph Claude can query, plus a second, separate graph that audits Cloze's help center for contradictions.

0
Brain nodes
0
Connections
0
ClozeKB nodes
0
Contradictions found
How it's built
Notes in, queryable graph out

graphify reads the source material, extracts concepts and their relationships, de-duplicates, and writes one graph.json that several consumers read.

📥 Sourcesgraph-visible material
Brain notes (Projects / Sessions / Stack) Articles + News digests 3,404 podcast summaries
🧪 Buildsupervised · Max-covered · $0 marginal
/graphify --update (you run it) concept + relationship extraction dedup pass (canonicalize)
🗄 graph.json9,724 nodes · 11,555 edges
brain-graph MCP (7 query tools) LLM wiki: 1,261 files, auto-regen Obsidian graph view
Two graphs
One brain, one audit

Same engine (graphify), two independent corpora, two MCP servers. Kept fully separate by design.

🧠 Brain graph

Your second brain across everything: notes, reading, listening.
Concepts (nodes)9,724
Connections (edges)11,555
LLM wiki files1,261
Served bybrain-graph MCP

📚 ClozeKB graph + audit

A pitch asset: Cloze's help center, graphed and fact-checked.
Concepts (nodes)2,572
Named communities291
Contradictions found41
Served bycloze-graph + cloze-audit
What Claude can do with it
7 query tools, offline & free

The graph is exposed as a read-only MCP server, so any Claude session can ask it questions. No tokens, no network, just the local graph.

get_node

Fetch a concept and everything known about it.

get_neighbors

What's directly connected to a concept.

query_graph

Search the whole graph by topic or term.

shortest_path

How two ideas connect, step by step.

get_community

Pull a whole cluster of related notes.

god_nodes

The most-connected hubs. Your core themes.

graph_stats

The size and shape of the graph at a glance.

+ 5 audit tools

cloze-audit surfaces contradictions, coverage gaps, and redundant articles with sourced quotes.

The design calls
Why it's built this way

Supervised ingest

You run /graphify --update by hand. It's Max-covered (no metered cost) and keeps a human in the loop, which matters because Brain ingests untrusted web text, an injection surface an unattended agent shouldn't touch.

Dedup pass

A canonicalization step merges duplicate concepts so the graph stays clean as new material lands.

Auto-regenerating wiki

An agent-crawlable wiki (1,261 files) rebuilds itself whenever the graph changes. Pure-local, no token cost.

Queries are free

Building the graph costs (supervised). Querying it doesn't. The MCP server reads the local file, so every Claude session gets the graph for $0.