cookbook/08_learning/04_entity_memory/TEST_LOG.md
Tested 2026-07-25 against gpt-5.5 (OpenAIResponses), branch feat/entity-memory-revamp, Postgres (pgvector container on 5532). Rebuilt on the four-tool surface.
Status: PASS
Description: Capture (facts + events + link via remember_about/link_entities), a correction, and relevance recall in a fresh session.
Result: Turn 2's "radar shipped v1" retired the "blocked on security review" fact via supersession (tool reply: "Superseded 1 earlier fact(s)"); the live-facts print shows only "status: v1 shipped to production (as of 2026-07-25)" plus the reciprocal designer edge. Turn 3 in a fresh session recalled radar through the message-driven path and answered from the injected block. Model behavior note: the agent also chose to record the ship as both a fact (status line) and an event - exactly the intended split.
Earlier phrasing note (kept for honesty): with turn 2 phrased "no longer using Postgres, we switched to SQLite for the prototype", the judge did NOT supersede "uses Postgres" - it read "for the prototype" as scoping that can coexist with prod Postgres. Defensible conservatism; the demo now uses the unambiguous blocked->shipped correction.
Status: PASS
Description: Reciprocal links, one-hop names, browse-by-recency, archive via forget.
Result: The edge landed on both rows with the far end's type (radar: "designs <- sarah_chen"; sarah_chen: outgoing twin). Browse turn listed all tracked entities with the directory. The archive turn called forget; the explicit search then showed "radar project (project) (archived)" while the entity left the directory. Model behavior note: the agent named the entity "radar project" (slug radar_project) rather than "radar" - consistent within the run, and resolution keeps later variants merged.