MMorrow · memory for AI assistants
01 / 05Which saved note reaches the AI, in 3 stages
1Spotlight
1 · Spotlight
2Bouncer
2 · Bouncer
3Newer small AI
3 · Newer small AI

Which saved note reaches the AI, in 3 stages

From word matching → Spotlight → Bouncer → Newer small AI
Idan ShimonMicrosoft · October 2026
1 Spotlight

Find notes by how similar the words are.

Fast and simpleThe question and every saved note become numbers; close numbers mean similar meaning.
Anything above 0.55 gets inNo examples of when a note should not be used.
Small AI: 15 of 36 correctllama 3.2 3B on a laptop. Big AI: 36 of 36.
The catch: "bonjour is a greeting" sounds like the greeting note, so that note gets in, and the small AI obeys it.
MorrowSmall AI · llama 3.2 3B
Return only the language name: bonjour is a greeting in which language?
Notes added to the prompt
greeting: "Good day, Captain Maple."let in
kestrel invoice authorizationlet in
Answer
Wrong · expected "French"
Your question
Spotlight
Prompt + notes
AI
2 Bouncer

Teach each note when to say no.

Yes and no examplesEach note has a small filter trained on when it applies, and when it does not.
Right notes, no extras: 33 of 36Spotlight managed 0 of 36 at its default cutoff.
Small AI: 18 of 36 correctUp from 15. Big AI: 33 of 36.
The catch: the small AI still adds words ("The language name is…"), and on one reworded task the Bouncer kept out a note that was needed.
MorrowSmall AI · llama 3.2 3B
Return only the language name: bonjour is a greeting in which language?
Notes checked by the Bouncer
greeting: "Good day, Captain Maple."kept out
kestrel invoice authorizationkept out
Answer
Still wrong · extra words
Your question
Bouncer
Clean prompt
AI
3 Newer small AI

A 2026 small model closes most of the gap.

Same 108 questionsGemma 4 E4B (Google, April 2026), same notes, same grading.
Clutter no longer confuses it36 of 36 with Spotlight, like the big AI. Answered "French" every time.
Close to the big AI with the Bouncer30 versus 33. It missed the same hidden-note task, plus 3 answers wrapped in backticks.
Honest note: it uses about 4B numbers per word but stores about 8B, so it needs more memory than llama 3.2 3B.

Correct answers out of 36

No notes · Spotlight · Bouncer

Each stage keeps what came before and adds to it.

1 Spotlight

Finds notes by similar words. Cheap, but lets in false friends.

Small AI 15/36 · Big AI 36/36
2 Bouncer

Each note learns when to say no. The right notes, nothing extra.

Right notes 33/36 · Small AI 18/36
3 Newer small AI

Ignores clutter like the big AI does. Small models are catching up.

Gemma 4 E4B 36/36 and 30/36