Adahas

We’ve all been there when a specific song on the radio instantly transports you back to a summer road trip years ago. Or when the scent of a particular perfume reminds you of an old acquaintance. These occurrences highlight the beautiful, chaotic, and deeply dynamic nature of human memory.
But how does this process actually work, and how does it stack up against the way artificial intelligence “remembers”?

Let’s explore this through a story.

How human memories work

Imagine a Tuesday morning. You’re walking to work, thinking about the warm croissant at your favourite bakery on the main road. You step off the curb and – SCREECH – a motorcycle clips you. You’re knocked to the ground, shaken and scared. You hear the panicked shouts, the wail of an ambulance, feel the sting of gravel rash.

In your brain, trillions of neurons are connected in a vast, intricate web. Your “favourite bakery,” “going to work,” “motorcycles,” and “the main road” all have their own little neighbourhoods in this web. The accident doesn’t create a new, separate file folder labeled “Terrible Morning, Feb 23rd 2026.” Instead, it rewires the existing map. The connections between “bakery” and “motorcycle” are suddenly strengthened and linked to new sensations: the sound of the ambulance siren, the sight of your torn trousers, the pain in your knee, the face of the worried onlooker who helped you.

This is neuroplasticity. Your brain is constantly being rewritten by your experiences. A memory isn’t a fixed file. It’s our ability to traverse this dynamic, ever-shifting web and pull together pieces of spatial (the place), temporal (the time), and sensory information. It’s a living, compounded experience. Every time you recall that morning, the neurons involved fire together, and the connections are subtly updated. It is a continuous process of creation, not just playback.

How LLM memory works

An LLM “brain” is a static set of weights in a vast, multidimensional vector space. Think of it as a frozen lake. The patterns are all there, but nothing can change them.

• No Organic Memory: An LLM cannot organically remember anything new from a conversation. Its “weights” are frozen after its initial training.
• Context Caching: LLMs have context windows to “remember” past conversations and in 2026, these windows are millions of tokens long. Context Caching allows LLMs to keep information active across sessions. But it’s still just a long, digital receipt that’s more akin to reading a note from a scratchpad.
• External Storage (RAG): To get around this, a technique called Retrieval-Augmented Generation (RAG) can be used. Think of it as giving the LLM an external filing cabinet. A RAG system creates a searchable database of vector embeddings. If you later ask the LLM about something, the system would search this database. Your query, mathematically transformed, would need to be similar enough to retrieve the correct file and feed it back into the LLM’s temporary context window.

The Fundamental Difference: Spontaneity vs. Retrieval

This is where the difference between the living brain and the frozen LLM becomes crystal clear.

• Memory Decay
  Human memories are fluid. The strong link between “bakery” and “pain” will slowly fade if you don’t revisit it. Other memories like new breakfast spots, different commutes, will layer on top, blurring the old connection. In a RAG system, the digital file of “the accident” remains in perfect, pristine condition forever, until someone manually deletes it.
• Unexpected Insight
  Years later, you see a pumpkin. It makes you think of a pumpkin spice latte. That leads to Starbucks, which reminds you of the coffee you used to drink at your desk. That thought leads to your old office, which is near the bakery, and suddenly, the memory of the accident floods back. This is a chain of loose, creative associations and is called an involuntary autobiographical memory. The brain “pushed” the memory to you.
  In AI, memory is a “pull” system. For an RAG system to have the same “thought”, your query would have to be mathematically similar enough to trigger a search. It has no capacity for such a winding path of insight.
• The Act of Remembering
  Most crucially, every time you recall the accident, your brain reconstructs it. In that process, the neural pathways involved are subtly changed. The memory itself is re-consolidated. You might focus on a different detail this time, strengthening some connections while letting others weaken. It is a continuous, dynamic process. An LLM’s knowledge is frozen in time, a snapshot of its training data. When an AI access a memory via RAG, the “file” remains exactly as it was written. it is a librarian reading a book, and the book doesn’t change because it was read.

Conclusion: Living with Memory, Not Just Storing It

We have reached a point where AI can store every grain of sand on a beach, yet it still cannot feel the tide come in.

While “unfreezing” a model to allow it to learn in real-time poses massive safety risks (how do we keep an AI from “learning” bad habits or false information?), it remains the final frontier. Until we move from Retrieval to Rewriting, AI will remain a brilliant librarian in a frozen library – capable of finding any fact, but incapable of being moved by a single one.