Ryan Yu.

Prototype · Agent memory · 2026

Memory should form an impression, not file a fact.

Most AI memory stores flat facts. Taste is built from behavior, held with different levels of confidence, and changes over time.

A useful memory should weigh what someone does more than what they once said, show why it believes something, accept correction, and let old beliefs fade.

From a drawer of facts
to a living point of view.

01

Memory today

one flat fact per slot
Heat

mildnonehigh

latest signal wins
Acid

avoids

stated once, at setup
Weeknights

pasta

no count behind this
A drawer the user maintains by hand.
02

Living model

an impression, held with confidence

Reaches for heat on weeknights

84%
9 signals · reinforced 2 days ago

Wants lighter meals

38%
stated once · contradicted by orders

Into sourdough

14%
not reinforced in 23 days · fading
Preference, confidence, provenance, and decay.

Every belief stays accountable.

01

Behavior

What someone repeatedly does matters more than what they once said.

02

Confidence

Every impression carries a visible degree of certainty, not a binary label.

03

Provenance

The system can show the signals behind a belief and accept a correction.

04

Decay

What is not reinforced fades. Memory changes as the person does.

An agent can only act consistently when its memory can explain what it believes, how sure it is, and change when it is wrong.

A product experiment, not a production memory system.

Try the live prototype ↗