Your AI brain is a Time Machine

In my AI journey I've read many books to get inspiration for things that could be made (what problems to solve) as well as how to make them (how to implement memory of AI, how to give AI sight).

I discovered a book that talks about time, and how our brain is effectively a time machine. Your Brain is a Time Machine

If you just want the tl;dr:

1. The brain is a Time Machine that remembers the past in order to predict the future.
2. The brain is a machine that tells time.
3. The brain is a machine that creates the sense of time.
4. The brain allows us to mentally travel back and forth in time.

Think of the next section as me thinking out loud. I'd love feedback, debate, or pushback.

How do we experience time?

We humans don't have a dedicated sense organ for time. The brains of all animals, humans included, are better equipped to navigate, sense, represent, and understand space than time.

As we build AI, specifically agents, we have the opportunity to enhance this sense of time. Our interactions and memories can get a timestamp, and whenever we interact with our AIs, this information could make those memories more or less relevant depending on recency.

I also believe that as we understand our minds even more, and enhance them with AI-powered capabilities, we may be able to "move back in time" in a mental sense, or even experience time differently.

I've read research papers (neurology-focused) that talk about how certain people have faster reflexes. Maybe we can tweak ourselves to experience time differently. Vision augmentation (via real-time CGI) might be part of that picture.

Evolution

Memory did not evolve to allow us to reminisce about the past. The sole evolutionary function of memory is to allow animals to predict what will happen, when it will happen, and how to best respond when it does happen.

This applies most directly at the application layer of AI: giving personal AI agents relevant memory.

For instance, when training transformer-based LLMs, we could add timestamps to data during training and inference, helping with ranking and semantic organization.

Example event representation (using MongoDB's document model-ish):

{
  "_id": 12231231,
  "event": "I experienced a car crash, I was on my way to work.",
  "time_stamp": "2025-09-02T00:00:00Z"
}

The timestamp could help with retrieval, linking memories not just by semantic similarity, but also by temporal context (e.g., connecting events around life milestones).

Your memories fade

Spike-timing-dependent plasticity (STDP): adjusts connection strength depending on the relative timing of spikes, strengthening if they are correlated in time, weakening if not.

We can apply this principle to AI memory systems through temporal correlation strengthening and decay. Consider learning to travel efficiently: after flying weekly for work over 5 years, you develop strategies like avoiding checked bags to skip baggage claim. This memory connection strengthens because "preparing for flight" and "successful trip completion" repeatedly correlate in time.

But if you stop traveling for 5 years, that strategy may become obsolete—airlines might now guarantee bags arrive before passengers (hypothetically). Without temporal reinforcement, the connection should weaken, making room for more relevant information.

In AI systems, we could implement this by tracking both semantic similarity and temporal correlation strength. Frequently co-occurring events within time windows get stronger retrieval weights, while unused connections decay—keeping inference fast even as storage grows cheap.

Transformer models today typically have fixed weights after training. They don't have a built-in decay or plasticity mechanism like biological neurons, unless you fine-tune or apply continual learning methods.

Remember, this is just a brainstorm. If you want to discuss or question these ideas, feel free to reach out: ArturoNereu.

If you want to read more about the content I've used to make sense of AI, check this repo: AI-Study-Group.