Node types¶
Every node in a Hadron memory has a nodeType. The type controls how the node
behaves in search, what it's typically used for, and whether agents and the
portal show it by default. Pick the type that matches the node's role; you
can always change it later.
Reference¶
| Type | Semantics | Search behavior | Typical use | Examples |
|---|---|---|---|---|
info |
First-class knowledge content. | Returned by default search. | The default — any node holding domain knowledge an agent should reach for. | A how-to article, a customer profile, a project decision write-up. |
abstract |
A summary or TL;DR of related content. | Searched first; ranked above info. |
Short summaries that should match before the full content does. | A 3-line TL;DR pointing at a longer explainer; an executive summary of a long doc. |
reference |
A pointer to an external source. | Returned by default search. | URLs, papers, legislation, third-party docs you want to cite without copying. | A node whose content is a citation block for an arXiv paper or a court ruling. |
record |
A factual record of something that happened. | Hidden by default; opt-in to include. | Chat messages, session logs, user actions — high-volume content that bloats search. | An individual chat turn under chats:<id>:messages:*; an agent session log. |
system |
Configuration or runtime scaffolding for the agent. | Hidden from search. | Conversation designs, prompts, partials, and other agent-internal nodes the agent reads by URN, not by search. | conversations:onboarding:welcome; prompts:partials:metadata-spec. |
Picking between neighbors¶
A few choices come up often. Two pairs are easy to mix up.
info vs. abstract¶
Use abstract when the node is a short summary that points at fuller
content elsewhere and you want it to surface first. Use info for the
fuller content itself.
A useful pattern:
infonode: a 1500-word write-up of how a customer ran their pilot.abstractnode, sibling: a 3-line TL;DR of the pilot, with a link or reference back to theinfonode.
When an agent searches for "pilot results", it hits the abstract first. If the agent needs detail, it follows the link.
If you don't have a separate summary — the node is the content — use
info.
info vs. reference¶
Use reference when the node's value is a pointer to something external
(a URL, a paper, a statute) and the content is mostly a citation block. Use
info when the node holds knowledge written for the memory itself.
reference: "ISO 27001 — see iso.org/standard/27001. Most relevant clauses for our SOC2 prep are 8.x and 12.x."info: "Our SOC2 readiness plan, drafted 2026-01, covers ISO 27001 clauses 8.x and 12.x as follows…"
The reference node is the breadcrumb. The info node is the content.
Defaults and visibility¶
- New nodes default to
infounless you specify otherwise. recordnodes are hidden from default search to keep the volume of chat/session content from drowning out real knowledge. To include them, pass an explicit option to the search call (includeRecords: trueonh-find-nodes).systemnodes are always hidden from search. The agent reaches them by URN (e.g.conversations:strategy:diagnose) when the runtime needs them.
Related¶
- Building an agent — where node types appear in the agent setup flow.
- Building a chatbot agent —
conversation designs and prompts use
systemnodes. - Mustache template syntax — variable resolution in node content.