TypeSafe’s cookbook wraps Jev in a small customer-memory system. It stores earlier tickets that contain lasting facts, then adds them to later support-routing requests for the same customer.

Separate the answer from the memory gate

  • A Choice question routes the ticket. A second, sequential Noul question asks whether the original ticket states a lasting customer fact.
  • The gate sees Jev’s answer as context but is instructed not to treat facts appearing only in that answer as evidence. Saved entries still include the answer, so an error can carry forward.
  • The example uses a Python dictionary keyed by customer ID. Each later request includes every saved entry for that customer.

A small demonstration, not a production benchmark

TypeSafe reports that its pinned Jev model routed three of six fictional follow-up tickets to the expected team without memory and all six with memory. The examples were written for the cookbook; the result was not independently reproduced and does not establish production accuracy.

The dictionary disappears when the process exits, grows without a limit and has no relevance filter or expiry. A real system needs access controls, retention and correction rules, plus a threshold evaluated on representative labeled cases. A typed decision does not guarantee that a memory or route is correct.