Finding
The PostgreSQL 19 / SQL-PGQ conv-26 smoke on released RememberStack v0.5.0 (050f9f61) passed direct live traversal but scored 0/2 on the selected LoCoMo category-1 questions.
- all 19 session versions reported pipeline, P1, live-graph, and P3 ready;
- the property-graph catalog was clean;
- a real two-hop neighborhood returned 8 nodes / 8 edges without truncation;
- a pair with no direct edge returned one shortest path of length 2;
- one category-1 answer used
answer_context and was too broad;
- the other called only
resolve_entity("Caroline"), received 92 exact current candidates with that name, and answered Unknown;
- a direct
answer_context call for the identity question did contain the load-bearing identity term.
This is upstream retrieval/entity-resolution quality, not a PostgreSQL 19, SQL/PGQ, catalog, or bounded-traversal failure.
Narrow follow-up
- Treat
resolve_entity as an identity-preparation call, not content-bearing evidence; the answer loop must not finish Unknown after only resolution.
- Investigate why one 19-session conversation produced 92 current entities named Caroline and define the intended cross-session identity/clustering lifecycle.
- Re-run the two category-1 items after the reviewed correction and record tool traces.
Safety boundary
Do not globally merge entities merely because their normalized names match. Same-name people are valid and the current resolver deliberately treats exact names as candidate generation rather than an identity verdict. Any identity correction needs provenance/scope and tests for distinct same-name entities.
Evidence
UMC PR #326 records the exact release digests, deployment, readiness, traversal results, score, final verified GCS receipt, and retained-volume state.
Finding
The PostgreSQL 19 / SQL-PGQ
conv-26smoke on released RememberStack v0.5.0 (050f9f61) passed direct live traversal but scored 0/2 on the selected LoCoMo category-1 questions.answer_contextand was too broad;resolve_entity("Caroline"), received 92 exact current candidates with that name, and answeredUnknown;answer_contextcall for the identity question did contain the load-bearing identity term.This is upstream retrieval/entity-resolution quality, not a PostgreSQL 19, SQL/PGQ, catalog, or bounded-traversal failure.
Narrow follow-up
resolve_entityas an identity-preparation call, not content-bearing evidence; the answer loop must not finishUnknownafter only resolution.Safety boundary
Do not globally merge entities merely because their normalized names match. Same-name people are valid and the current resolver deliberately treats exact names as candidate generation rather than an identity verdict. Any identity correction needs provenance/scope and tests for distinct same-name entities.
Evidence
UMC PR #326 records the exact release digests, deployment, readiness, traversal results, score, final verified GCS receipt, and retained-volume state.