The least calibrated probability at which a candidate note is kept.
Set from the memory-note gate measurement (plan Task 3.4, 2026-09-29): 60 synthetic conversations,
337 candidate notes labelled durable, ephemeral, wrong or speculative, two repeats. At a calibrated
0.6, Jev kept durable notes at 96.1% precision and 95.4% recall, and kept 6.2% of the
conversation-only (ephemeral) notes. A proxy self-report scorer's confidence >= 80 kept 40.0% of
them at 85.3% precision and 98.7% recall. Neither kept a wrong note; the gate kept 3.4% of the
speculative ones, the proxy none.
The baseline is a proxy, not today's Memory Manager. Its self-reported confidence came from a
separate scoring call with a four-band rubric (90-100 stated, 75-89 strongly implied, 50-74
possible, below 50 skip). Production's Memory Manager - Extract Notes template has three bands
(90-100, 80-89, below 80 skip), and skip rules the proxy doesn't have: one-time requests, matters
resolved within the conversation, and task-scoped clarifications. Those rules drop exactly the
ephemeral notes counted here before any confidence filter runs, so production probably keeps fewer
than 40% of them. The gate's own figures stand; the comparison is against the proxy.
The least calibrated probability at which a candidate note is kept.
Set from the memory-note gate measurement (plan Task 3.4, 2026-09-29): 60 synthetic conversations, 337 candidate notes labelled durable, ephemeral, wrong or speculative, two repeats. At a calibrated 0.6, Jev kept durable notes at 96.1% precision and 95.4% recall, and kept 6.2% of the conversation-only (ephemeral) notes. A proxy self-report scorer's
confidence >= 80kept 40.0% of them at 85.3% precision and 98.7% recall. Neither kept a wrong note; the gate kept 3.4% of the speculative ones, the proxy none.The baseline is a proxy, not today's Memory Manager. Its self-reported confidence came from a separate scoring call with a four-band rubric (90-100 stated, 75-89 strongly implied, 50-74 possible, below 50 skip). Production's
Memory Manager - Extract Notestemplate has three bands (90-100, 80-89, below 80 skip), and skip rules the proxy doesn't have: one-time requests, matters resolved within the conversation, and task-scoped clarifications. Those rules drop exactly the ephemeral notes counted here before any confidence filter runs, so production probably keeps fewer than 40% of them. The gate's own figures stand; the comparison is against the proxy.