What ontology means here
A small vocabulary of record types and valid relationships. It gives structure to a memory without requiring a broad automatic knowledge graph, an LLM, or a separate graph database. A vector can suggest similarity; an explicit relationship records what a note actually refers to.
The initial types
Asset, AssetRevision, Region, TextObservation, Annotation, Collection, Tag, and Source. Optional user-defined subjects can follow actual needs. Schema definitions and revisions should be explicit.
The useful relationships
Annotation → refers_to → Region
Region → part_of → AssetRevision
AssetRevision → revision_of → Asset
Asset → member_of → Collection
Asset → captured_from → Source
Annotation → quotes → TextObservationWhy it helps an assistant
“Make this button quieter” can point to one rectangle, its quoted label, your note, and the exact screenshot revision. Retrieval can fetch that region and inspect its source instead of handing over an unrelated image with similar colors.
Keep assertions grounded
Each relationship retains provenance and confirmation state. A user-created link, a source fact, and a proposed model association are different evidence. A place in OCR is not proof of capture location. Rejected or corrected associations must not keep influencing retrieval.
What belongs later
Automatic entity extraction and broad relationship inference are later experiments. Compare them against explicit links and hybrid search, including false links and correction effort, before adding them to the default workflow.