Memory Write Gate
Decides whether what the user just said should become long-term memory, so the assistant's memory holds lasting facts and not noise. Jev answers four questions in one call: is it durable (still true and useful in a month), is it about the user themself, is it already known (given the stored memories; a change to a stored fact counts as new), and does it contain a secret? Code stores it only if all four say so; secret-shaped strings (project 12's detectors) are refused before any call. Against it: storing everything, and an LLM asked "should I remember this?". Measured: accuracy, junk stored, facts lost.
With Jev
◐ ▮▮▮ ▦Without Jev
LLM · CODECalls Jev and every baseline for real. Nothing is saved.
Runs every labeled row, one at a time: 22 rows × every variant, all real, paid calls. Rows run sequentially so latency measures the model, not a traffic jam. The run is saved as it goes and lands in History.