How the Entor engine traverses connected memory
Connected nodeAction 01
- Current
- policy.v4
- Relationship
- superseded by
- Candidate
- amendment
- Recurrent memory
- Carried from previous step
Awaiting Transformer output
Traverse all neighbors?Yes
Full read?No
Memory updatedReady for next step
02 / Engine
Relationship
Every decision starts with two connected nodes.
The engine begins with a current node, one relationship, and the candidate on the other side.
- Traverse
- Full read
- Ignore
03 / Performance
Locked synthetic testMeasured retrieval quality
Ranking the right paths. Returning the right context.
The locked benchmark reports graph-ranking and selected-context quality using the metrics each task actually measures.
- 1.000
- Relationship traversal MRR
- 0.99875
- Bounded-context F1
- 0.99966
- Locked combined score
The correct relationship ranked first on the locked test.
Selected context quality at top 8 / 1,200 tokens.
Balanced score across the benchmark objectives.
Locked synthetic test · 160 cases · 64 candidates per case · Qwen-teacher-derived corpus. Not customer or production validation.