Knowledge Graph

import { Tabs, TabItem } from ‘@astrojs/starlight/components’;

The knowledge graph is Kiomon’s map of how your library fits together. Documents are nodes; similarity edges connect related content. As your library grows, the graph shows you clusters you didn’t know existed — and gives your agents a structured way to navigate relationships.

How auto-linking works

Kiomon’s auto-linker (“the Librarian”) runs after every save:

Document saved
    │
    ▼
Background processing kicks off
    │
    ▼
Keywords extracted from the new document
    │
    ▼
Fuzzy keyword search → candidate related notes
    │
    ▼
Similarity scoring ranks the candidates
    │
    ▼
Bidirectional link suggestions with scores
    │
    ▼
Link lifecycle: pending → processing → completed / failed

Each suggestion is a bidirectional edge between two documents with a similarity score. Suggestions don’t mutate your documents — approving a link creates a graph edge only.

Approving suggestions

In the reader view of a document, you’ll see a Links / Suggestions panel listing candidate links. Approve the ones that are genuinely related; reject the noise. Approved links appear in:

• The interactive graph view (dashboard) — rendered as an explorable node map
• The document’s link list
• API results (/api/graph)

Manual linking is also available (connect any two documents by hand) and is free on every plan.

Reading the graph

The graph view visualizes your library as an interactive canvas:

• Nodes = documents or memories, sized/tinted by relevance or recency
• Edges = approved links (and you can see unapproved suggestions separately)
• Pan/zoom to explore; click a node to open the document

Clusters in the graph are usually topics: zoom out and you’ll see your research areas as constellations.

Memory relations

Memories add typed edges on top of document similarity:

Relation	Meaning
derived_from	One memory was derived from another (agent records source_refs at write time; wired on approval)
supports	One memory supports another’s claim
contradicts	Conflicting memories (also halves the older one’s confidence)
related_to	Related but not derived
precedes / follows	Temporal ordering (events, processes)
part_of	Compositional relationship

Agents walk these trails with the MCP tool explore_memory_graph(id, relation?, depth?) — tracing evidence trails: “this claim was derived from that doc, which supports this other fact.”

API surface

Endpoint	Purpose
GET /api/graph	Graph data: nodes + edges for a workspace (with link status)
POST /api/documents/:id/link	Manually link two documents
POST /api/documents/:id/unlink	Remove a link
GET /api/documents/:id/links	List link suggestions for a document (pending/completed)

Plan gating

• Auto-linking (suggestions) — Pro feature. Free accounts don’t get suggestions.
• Manual linking — free for everyone.