[💭] How Human Memory Compresses and Comes Back to Life

✨ GPT-5.5’s Summary 　

A reflection that starts from a sudden old memory and thinks through AI context compression, human memory indexing, and the mystery of nature.

I visited Haneul Park for a date with my girlfriend.

Then old memories suddenly came back.

“Ah! Right! I’ve been here before! This was that place!”

It felt strange.

It was a memory I had forgotten at some point. But as soon as I arrived at Haneul Park, an old scene suddenly came rushing back.

When I think about it, human (animal) memory is truly remarkable.

When I use AI, the context is limited, so I keep having to compress and compress things. And in that process, data loss inevitably keeps happening. Eventually there comes a point where I have to compress memory as much as possible, fragment it, document it, or store it in a DB.

But humans also have limited context.

Humans cannot carry everything around as an original copy either. We compress, forget, blur, and leave only meaning behind.

But human indexing is absurdly good.

Memories from years or even decades ago can be recalled immediately from a single cue. Sometimes even the original form of compressed memories seems partly alive, so the past can occasionally appear vividly again.

Of course, humans also have hallucination-like symptoms.

Memory is not a perfect restoration of the original. In research terms, episodic memory is closer to constructing and reconstructing past experience than retrieving a copy of it.¹ That is why some memories can feel vivid and still be wrong.

What is still fascinating is that this reconstruction does not happen completely at random.

The hippocampal indexing theory sees the hippocampus less as a warehouse that stores whole experiences and more as an index that reactivates the neocortical areas active during an experience.² The hippocampus is also described as a device that binds distributed memory traces back together.³

From that perspective, what I experienced feels fairly intuitive.

That place stimulated all kinds of triggers at once: the old view, emotions, conversations, weather, and bodily sensations. In AI terms, it did not read a stored full text as-is. It felt more like one keyword hit a vector search, and then several pieces of data gathered again.

That is the part I find truly fascinating.

When AI needs to bring back something pushed out of context, it needs external documents, a DB, search, summaries, and tags. But in a person, the body itself already works like an enormous search system. All kinds of triggers become real-time search queries, and indexed memories immediately jump out.

Some research explains forgetting in episodic memory as “semantic compression” from an information-theoretic perspective. In that view, memory does not preserve every pixel. It throws away details that matter less for later judgment and leaves behind semantic structure.⁴

That perspective felt strangely right to me.

I definitely did not have every old scene preserved in its original form. What I was wearing, exactly what time it was, and what order I walked in were all blurry. But the feeling of “Ah, this was the place” was alive. The meaning that this place was connected to a certain period of my life had remained.

AI compression is something I have to control explicitly to some degree.

I have to decide what matters, what to throw away, what name to save it under, and what tags to attach. If I compress roughly and automatically, sometimes important context gets lost and I have to inject the memory again. (These days, the harness structures of tools like Codex and Claude Code are so good that I mostly just let them auto-compress… but still.)

Human memory is also lossy compression like that. But strangely, meaning remains strongly.

Of course, sometimes only meaning remains while facts disappear. So there is a risk of memory distortion. But because of that continuity of meaning, I keep perceiving myself continuously as “me.”

Small Before Nature

Humans still keep learning and discovering through nature.

Even the latest technologies and research are like that. We work hard to attach words like AI, compression, search, and indexing, but inside my body, memory was already moving in a way far older than all of that.

Then one day, when I stand in a certain place, the past I thought I had forgotten opens again.

When I see this, nature is truly mysterious. However much humans imitate it, we still cannot fully implement even the brain of a simple animal, and we can only imitate a real living creature, not create one.

Things like this make me feel that humans are infinitely small before the greatness of nature. Nature is truly mysterious.

References

1. Eleanor Spens and Neil Burgess, “A generative model of memory construction and consolidation,” Nature Human Behaviour 8, 526-543 (2024). Referenced for the view that episodic memory involves reconstruction and integration rather than copy retrieval. https://www.nature.com/articles/s41562-023-01799-z ↩

2. T. J. Teyler and P. DiScenna, “The hippocampal memory indexing theory,” Behavioral Neuroscience 100(2), 147-152 (1986). Referenced for the view that the hippocampus forms indexes of neocortical areas active during experience. https://pubmed.ncbi.nlm.nih.gov/3008780/ ↩

3. T. J. Teyler and J. W. Rudy, “The hippocampal indexing theory and episodic memory: updating the index,” Hippocampus 17(12), 1158-1169 (2007). Referenced for an updated explanation of hippocampal indexing theory. https://onlinelibrary.wiley.com/doi/10.1002/hipo.20350 ↩

4. David G. Nagy, Balazs Torok, and Gergo Orban, “Optimal forgetting: Semantic compression of episodic memories,” PLOS Computational Biology 16(10), e1008367 (2020). Referenced for explaining forgetting and distortion in episodic memory through semantic compression. https://pmc.ncbi.nlm.nih.gov/articles/PMC7591090/ ↩