Relationship graphs

Store structured facts as entity-relationship triples the Agent can traverse.

Some memory is naturally graph-shaped: Alice → works at → Acme, Order 42 → refunded on → 2026-07-15. Storing these as relationship triples lets the Agent traverse — following relationships to answer structural questions.

Where relationship graphs fit

You can store this info as free-text notes and rely on retrieval to find them. That works. But relationship graphs are better when:

• The Agent needs to answer multi-hop questions (“who else works at the same company as Alice?”).
• You have canonical entities with stable IDs (customers, products, tickets, projects).
• You need to filter or aggregate over relationships (“how many refunds did we issue this quarter?”).

Structure

A triple is (subject, relationship, object). Both subject and object are entities with IDs. The relationship is a named edge type.

(customer:alice) --[works_at]--> (org:acme)
(order:42)       --[refunded_on]--> (date:2026-07-15)
(order:42)       --[belongs_to]--> (customer:alice)

Entities can have attributes (name, created_at, custom fields) beyond just an ID.

Creating a graph memory space

Choose Index type when creating a memory space, and turn on Graph mode. This enables the entity/relationship model on top of the standard vector index.

Writing triples

From an Agent:

• Use the remember_fact tool with subject, relationship, object.
• Or turn on Auto-extract triples on the space — Nora extracts entities and relationships from freeform Agent notes automatically.

From the CLI:

nora memory triple <space> \
  --subject "customer:alice" \
  --relationship "works_at" \
  --object "org:acme"

Traversing

The Agent gets two tools when attached to a graph space:

• recall_facts — retrieve facts about an entity by ID or name.
• traverse — follow edges. Ask “what does Alice work at?” and get the answer as a set of (entity, path) pairs.

Both tools are automatically added when a graph space is attached.

Bidirectional edges

Relationships can be marked symmetric (same_as, sibling_of) so traversal works both ways. Or explicitly directed (works_at, parent_of) with a defined inverse (employs, child_of).

Configure per relationship type in the space schema.

Entity resolution

Multiple mentions of the same entity should resolve to the same node. Nora auto-resolves on:

• Exact ID match.
• Alias match (list aliases per entity).
• Fuzzy name match above a similarity threshold (adjustable).

Manual merge available: Space → Entities → select duplicates → Merge.

Contradictions

If a new triple contradicts an existing one (works_at says Acme, new triple says Beta), Nora doesn’t silently overwrite. It flags the contradiction for review or, if you turn on auto-supersede on conflict, the newer triple wins and the older becomes historical.

Graph vs. causal graph

Relationship graph memory is entity-relationship data the Agent accumulates. A causal graph (see Causal Graph overview) is a domain-authored graph of cause-and-effect variables, used for verification and analysis. Different tools, different jobs — you can use both.