Memory in Artificial Intelligence Part I: Taking Inspiration from the Human Brain
Memory is one of the most prominent elements of human cognition and one that is becoming increasingly important for artificial intelligence(AI) systems. As AI agents evolve and tackle more complex scenarios, the role of memory should become more relevant. Currently, memory models are one of the most notably missing components of AI platforms and frameworks.
In the past, I’ve written about the role of memory in AI solutions. However, the more I think about this, the more I realize that architecting efficient memory systems for AI agents is ridiculously complicated. Maybe the best way to get an idea of this complexity is to draw inspiration for the human brain.
The Complexity of Human Memory
The creation of memories in the human brain involves very complex cognitive processes. When receiving sensory information from cognitive channels such as vision, touch, language, taste, etc the brain stem routes the information to the thalamus which decomposes it and sends it to different labels of the brain for evaluation. The result of that evaluation is what we know as SHORT-TERM-MEMORIES which are form in our consciousness and last for short periods of up to a few minutes.
Here is where things get more complex. In order to capture memories for a longer duration, the information has to flow through the hippocampus where is partitioned and sent to various cortices of the brain. This means that LONG-TERM-MEMORIES are not stored as sequences but decomposed as individual fragments and scattered throughout the brain. For instance, emotional memories are typically stored in the amygdale while words are captured in the temporal lobe. Mysteriously, the brain is able to maintain the association between all these fragments and the entire memory can be reconstructed instantly based on related cognitive experiences. Let’s take, for instance, the memory of going to a concert. The brain will decompose that memory into thousands of categories such as music, colors or emotions and store them across different sections of the brain. Time after, a single cognitive experience related to that memory -such as listening to a melody from the same band- can make the brain reconstruct the entire memory of the concert. How exactly does this process takes place is largely a mystery known in neuroscience as the “binding problem”.
The Most Fascinating Thing About Human Memory
We typically associate memories with the past but, in fact, they are typically recalled when we are thinking about the future. This is, in my opinion, the most fascinating aspect of human memory. A famous neuroscientist once said “the purpose of memory is to predict the future”. I know it sounds counter-intuitive at first but it makes perfect sense. Cognitive research has shown that the areas of the brain used to store long term memories are actively stimulated when we are immersed in future-related activities such as planning a trip or making a strategic business decision.
How Does This Relates to AI?
I hope the previous examples gave you an idea of the complexities of the human cognitive processes related to the formation and retrieval of memories. To be efficient, AI agents will need to leverage hierarchical, federated rapid storage systems that can be used to “create memories” from the cognitive data source such as vision, voice, language, etc. What should be the capabilities of AI-memory systems? We will speculate about that in a future post.