THE STRANGE SIMILARITY BETWEEN MEMORY AND AI

Think of a red lemon. And now blue elephants walking across mountains. Can you store those ideas as neat little files in your memory?

Turns out that’s not possible. Not because the concepts are somehow resistant to storage and playback. But because storage and playback isn’t how our memories work.

Science has known for a century that memory is a process of reconstruction rather than recall. And now, following the publication of a new study, it has a far more detailed grasp of how those thoughts unfold. From a compressed state, into our conscious mind.

At the centre of the process sit ripples - brief bursts of electrical activity. A ripple appears in the hippocampus, which connects it to the cortex. Which then expands into a richer, more differentiated pattern.

This is where red is separated from blue - and elephants in the mountains from others in your home office. And what blossoms into a memory.

The greater the expansion, the faster the memory appears. And the stronger the original association that returns to your mind.

Plus, it would seem, similar machinery is used to help us imagine things. It's part of what allows us to bind familiar, recallable elements together, into scenes that have never happened. Like, we presume, blue elephants in your home office.

So why, in the context of digital intelligence, do we care?

While the hypegeist missed it, the study reminds us of instructive ways of thinking about digital intelligence. Highlighting some similarities - and important differences - with human cognition.

In an LLM, the prompt is the cue. It activates relationships which, having been compressed into the model’s weights, are unfolded by successive layers. Which leads to context-sensitive representations.

The compression and unfolding work very differently in practice. But they share a conceptual shape with our own brain.

Both systems can also produce what the original cue doesn't contain. In humans, reconstruction supports imagination. In an LLM, generation can produce a useful inference - or, when the output outruns the evidence, a hallucination.

It also reminds us that while human recall can change the person, today’s digital intelligence remains fixed during inference. The context can influence the next answer, but that adaptation disappears when the session ends.

Continuity, learning and consequence therefore belong to the surrounding system - and only when engineers build them in.

So what’s today’s in-the-end-at-the-end?

You might be discombobulated to know exactly how complex the process is when you think about a red lemon. But at least you understand a little more about how those blue elephants showed up in your home office.