CHOOSEAgents & automation
Jev Agent Memory Test
Sean VanWinkle@SeanV6790
Jev successfully passed a rigorous agent memory test involving a large number of confusable memories, high-option questions, and mixed question types, demonstrating its ability to handle complex state and reasoning.
Tried a hard agent memory test on Jev that I honestly thought it might fail.
80 similar memories, 30 mixed questions (choice, noul, score), high option counts, and a lot of near misses plus questions where the right answer is none.
A normal reranker cant do that mix. Most other decision style models cant hold that much state at once. It got the hard referent and absence questions right and was only a bit softer on the score ratings.
Whole pass took 302ms.Not a toy demo. Just a case where the size plus mixed question types actually matters.
The test was 25k tokens and covered
large state (80 detailed, highly confusable memories),
high cardinality (81 options on most Choice questions)
adversarial instructions that forbid topical leakage
multi-hop / temporal / status-sensitive reasoning
explicit absence probes (NONE)
mixed primitives (Choice + Noul + Score) in one pass
context pressure near the practical limit
Jev passed every single question.
To see the test set/run
console.typesafe.ai/playground
