Memory graph and time travel
How entities, relationships and decisions are derived from your items, why connections wait for your approval, and what reading the graph as of a date means.
Breno does not only store text. It builds a graph of the things your material talks about, and keeps it honest.
Entities and relationships#
Reading an item, Breno recognizes entities (people, projects, products, organizations and other things) and the relationships between them. Each is stored with the passage that supports it, so a connection always points back to evidence.
- Entities and relationships are stored in the database and projected to a graph store used for retrieval.
- Purging an item removes the entities and relationships that only it supported.
- When an item is processed again, the previous extraction is cleared before the new one is written, so a re-run replaces rather than adds.
You approve what changes#
Breno proposes relationships and possible duplicate entities. They wait in Memory, Review. Nothing changes in the graph until you approve it, and a rejected proposal is dismissed. This keeps the graph something you can trust: it contains what a person agreed with.
Decisions#
A decision is a record with its reason, its status (active, superseded or replaced), who made it and when, what else was considered, and a link to its source. When a decision is replaced, the old one stays, marked as replaced and pointing at what replaced it. See Memory and decisions.
Time travel#
Because memory is kept as history rather than overwritten, the graph can be read as it stood on a date. In the graph, choose a past date and you see the relationships that were known then. It counts relationships as of that date: it answers "what did we know, then?", which is different from "what is true now?".
Confidence#
Entities, relationships and records carry a confidence value computed from several signals (the model's own confidence, how complete the extraction is, how well it validated, and the quality of its evidence). Low-confidence items are flagged, not hidden, so you see uncertain knowledge with its uncertainty.
Access#
The graph follows the shared read rule: results are filtered to items you can read. See How access works and the caveat in Known limits.