How Morrow Decides Which Memory Note Reaches the AI

How Morrow Decides Which Memory Note Reaches the AI A workflow diagram generated by Archify. 01 / Path A · Word-Similarity Spotlight 02 / Path B · 769-Number Learned Bouncer 03 / How Different AIs React 1. Read Question 2. Test the Saved Note 3. Build Brief & Answer Saved Note Only · no negative examples · Path A · Word-Similarity Spotlight › 2. Test the Saved Note Saved Note Only no negative examples Cosine Spotlight · 0.692 >= 0.55 (passes) · Path A · Word-Similarity Spotlight › 2. Test the Saved Note Cosine Spotlight 0.692 >= 0.55 (passes) Stuffed Brief · 75 extra rules in 36 runs · Path A · Word-Similarity Spotlight › 3. Build Brief & Answer Stuffed Brief 75 extra rules in 36 runs Giant Cloud AI · 36/36 (ignores clutter) · Path A · Word-Similarity Spotlight › 3. Build Brief & Answer Giant Cloud AI 36/36 (ignores clutter) Your Question · bonjour is a greeting... · Path B · 769-Number Learned Bouncer › 1. Read Question Your Question bonjour is a greeting... Meaning Scanner · 768 numbers (frozen) · Path B · 769-Number Learned Bouncer › 1. Read Question Meaning Scanner 768 numbers (frozen) Yes + No Examples · 3 use here · 4 not here · Path B · 769-Number Learned Bouncer › 2. Test the Saved Note Yes + No Examples 3 use here · 4 not here 769-Number Bouncer · 0.245 < 0.50 (blocks) · Path B · 769-Number Learned Bouncer › 2. Test the Saved Note 769-Number Bouncer 0.245 < 0.50 (blocks) Clean Brief · 33/36 exact note sets · Path B · 769-Number Learned Bouncer › 3. Build Brief & Answer Clean Brief 33/36 exact note sets Small Local AI (3B) · 18/36 clean vs 15/36 noisy · How Different AIs React › 3. Build Brief & Answer Small Local AI (3B) 18/36 clean vs 15/36 noisy Saved Turn & Receipt · pause or delete anytime · How Different AIs React › 3. Build Brief & Answer Saved Turn & Receipt pause or delete anytime word match admits noise 100% score draws line 0 false alarms focused Legend User UI Agent logic Policy Tool action Context / trace Cloud service External system

The Sticky-Note Bouncer

  • • Every saved rule stays in plain English inside SQLite.
  • • Only a tiny 769-number gatekeeper learns when to let the note through.

Why False Friends Trick Search

  • • Standard search only knows positive examples, so 'greeting' in a French question triggers Captain Maple.
  • • Negative examples ('Never use here') draw a wall between 'Greet me' and 'What is a greeting?'

Small Brain vs. Giant Brain

  • • Small local models get distracted when extra notes clutter the brief (15/36 -> 18/36 with the bouncer).
  • • Giant models shrug off clutter (36/36), so hiding a needed note by accident hurts more than extra noise (33/36).