Associatron Demo: Cue → Competition → Spark → Recall

 When we see a car from 40 years ago, an Associatron recalls the family and episodes from that time. But today’s LLMs would tell us used-car information instead.

Most of the time, those memories remain forgotten.

Yet the moment you see a gymnasium — and catch the smell of Air Salonpas — you can vividly remember someone from 40 years ago.
Rather than trying to “fill the gap” between these two kinds of intelligence, I’m exploring how they can be shared.

This is a single-page demo of Nakano’s Associatron-style associative memory, designed to visualize something that modern “AI memory” demos often hide

Recall is not a database lookup.
Recall is a competition between memories.
In everyday life, a memory is rarely retrieved by a perfect key.
Instead, a vague cue—smell, atmosphere, partial shapes—activates multiple candidate memories at once.

Those candidates compete, and one basin finally wins.

A cue is rarely perfect.

A smell, a mood, a vague shape—

and suddenly a whole episode returns.

This demo makes that process visible as

Cue (smell + extra conditions) → Top-2 rooms → memory competition → Spark (overlap) → Recall

Why “rooms”?

In this demo, each memory pattern is stored inside a conceptual room.
A room is a container of stored traces (patterns)
Similar cues tend to open similar rooms

When a cue arrives, the system does not immediately pick one correct memory
→ it first narrows down to Top-2 rooms, where recall is most likely to occur

This is important because it matches human recall:

we do not search the entire brain uniformly—we recall within a contextual neighborhood.

Cue input: ambiguity is the point

You can draw a cue (a rough sketch), and optionally inject noise.
This is not a bug—it is the essence of associative recall.

In real memory recall:
●the cue is incomplete
●the cue is noisy
●multiple episodes are partially activated

Therefore the recall is not a clean match, but a dynamical process.
Spark (overlap): where memory becomes visible

After competition starts, the demo displays the “spark” effect:
Overlapping features among the winning memories light up.

This is the key visualization:

not “the final answer”

but the shared structure that caused convergence

In other words, the spark shows why a certain basin wins.This is very different from classifier-like AI demos that only display a final output label.

Recall: reconstruction from the winning basin
Finally, recall is performed using the winner memory (across rooms).

The resulting image is not “replayed pixel-by-pixel.”

Instead, recall is a reconstruction driven by the associative dynamics—

exactly as Nakano’s Associatron was meant to demonstrate.

Associatron vs modern AI: the difference
Modern AI systems are optimized for correct answers.
But associative memory is not about correctness.

It is about:

●attractors (basins)
●context
●competition
●overlap
●recall dynamics

The goal is not to find the correct memory, but to reproduce the phenomenon of “something comes back to mind.”

This demo is an attempt to make that process visible.