Entities, evidenced edges, merges and the canvas explorer.
The relationship graph is a mind's structured picture of the people, systems, projects, places and things it has encountered, with typed, evidenced connections between them. It is built from what the mind perceives and corrected by operators; it is never populated from a model's imagination without an event behind each edge.
An entity has a kind (person, system, service, host, project, organisation, place, document, concept and a few more), a canonical name, aliases, attributes, mention counts and first and last seen times. Deterministic extraction finds candidate entities in observations and conversation (capitalised names, hostnames, repository names, quoted titles) and links each mention to the event it came from. Entities link to the memories, beliefs and events that mention them.
Edges carry a relation from one shared vocabulary, a confidence and evidence references to events, and are weighted by how often and how credibly they were observed. The vocabulary keeps association and causation apart. co_occurs_with and temporally_precedes record what was seen together or in order; co-mention within one observation writes nothing stronger. possibly_caused_by is a hypothesis: the association process proposes it for notable events close in time, always as a hypothesis belief with the alternatives "coincidence" and "a third factor caused both", and a prediction that the effect should recur. caused_by is reserved for established causes stated by an operator or recorded as a correction; nothing automatic writes it. asserted_by links a proposition to the person who put it forward. An explicit statement ("db-primary is the new name for server-3") still creates a strong edge of the kind it states.
Operators can merge two entities that turned out to be one (the merged entity becomes an alias, its edges and links move, and the merge is an event the mind remembers), or mark two as distinct so extraction does not merge them again. Both are audited.
The graph page renders a bounded neighbourhood around a chosen entity on a canvas with a force layout, with edge thickness for weight and colour for kind. It is deliberately capped in size; the detail page for an entity lists everything without the cap.
Each mind keeps its own entity graph (entities, entity_relationships, entity_links), mind-scoped like everything else.
Kinds: person, system, project, document, concept, repository, organisation, place, other. Names are canonicalised (lower-cased, trimmed) for matching; aliases accumulate through merges. Every mention bumps mention_count and last_seen.
Extraction is deterministic and conservative: hostnames (atlas.local), hyphenated identifiers (server-09, web-api-prod), letter-digit names and capitalised words become entities; plain lower-case nouns do not. The perceive prompt may add entities with kinds when a model is configured. Operators can correct the kind and description of any entity.
Edges carry a relation from the shared vocabulary (co_occurs_with, related_to, discussed_with, temporally_precedes, possibly_caused_by, caused_by, supports, contradicts, depends_on, belongs_to, observed_during, derived_from, similar_to, follows, asserted_by), a confidence, a source, a count of reinforcements and the event that produced them. Co-mention in one message yields discussed_with (speaker → entity) and co_occurs_with (entity ↔ entity) at low confidence; repeats raise the count but never change the kind. Operators can add typed edges with higher confidence.
Repeated association is evidence of association, nothing more. The vocabulary keeps the distinction explicit:
co_occurs_with and temporally_precedes record what was observed together or in order;possibly_caused_by is a hypothesis: the association process proposes it for notable events close in time, always as a hypothesis belief with the alternatives "coincidence" and "a third factor caused both", and a prediction that the effect should recur after the cause;caused_by is reserved for established causes: an explicit operator statement or a correction. Nothing automatic writes it.Speech acts are not world events: questions, hypotheticals, requests and consecutive turns by one speaker never seed a causal hypothesis.
entity_links record which memories, beliefs, events, goals and predictions involve an entity, so the entity page answers "where has this come up".
Merge re-points edges and links to the survivor, adds the loser's name as an alias, and keeps the loser row with merged_into so old references still resolve. An operator_action event and an entities_same correction are recorded.entities_different correction and event are recorded; it documents the operator's judgement and is visible to later extraction./minds/{mind}/graph renders a bounded subgraph (node cap, kind filter, optional focus entity with its neighbourhood) on a canvas with a small force-directed layout written in plain JavaScript. Clicking a node loads its panel (edges, links, corrections); double-clicking focuses on it. The API endpoint GET /api/v1/minds/{mind}/graph returns the same nodes and edges as JSON.