Human and AI Memory: Like Peas in a (Somewhat Different) Pod


cauri jaye

on •

Jul 27, 2023

Have you ever wondered if your brain and that fancy AI language model you use have something in common? Spoiler alert: they do! Although human and AI memory systems are like apples and oranges in many ways, they share some intriguing similarities when it comes to storing and recalling information. Let's dive into this memory mash-up as we explore the shared traits of biases, distortions, grouping, summarising, and reflection.

Two new heroes of AI memory include vector databases and the new summarising techniques from LLMs.

Vector Databases: The Superhero of AI Memory

Picture vector databases as the superhero of AI memory, swooping in to save the day with their high-dimensional vector spaces. Each piece of information gets its own special spot in this space, and similarities between points are measured with fancy-sounding metrics like cosine similarity or Euclidean distance. The real superpower of vector databases? Their ability to tackle huge amounts of data with ease and perform lightning-fast similarity searches. This technique really shines in applications like natural language processing, image recognition, and recommendation systems.

Summarising: Squish it! The Art of Bite-sized AI Wisdom

Enter summarising, another star player in the AI memory game. It's all about squishing large volumes of information into neat, bite-sized representations. With the help of AI models like transformers and recurrent neural networks, machines can generate summaries of text or other data formats, making it a breeze to extract essential info. Summarising lets AI systems focus on the juiciest parts of the input data, lightening their cognitive load and boosting overall performance. It's like having a personal information organiser, making life just that little bit easier.

Memory Selection: Picky Humans and Picky AI

When it comes to storing memories, both humans and AI can be quite choosy. Human memory is a complex and multidimensional process involving encoding, storage, and retrieval of information. Imagine our brain as a picky eater, choosing which memories to savour and which to discard. Factors like personality, genetics, upbringing, and psychology impact our memory "diet."

The process of consolidation in human memory, which involves the hippocampus and other brain regions, aids in the transfer of information from short-term to long-term memory. Memory selection in large language model-based AI systems, like GPT, on the other hand, takes cues from the design of the algorithms, the training data used, and the instructions provided for specific tasks such as summarising.

Parallels Between Summarising in AI and Reflection in Human Memory

There are striking similarities between the process of summarising in AI and reflection in human memory:

  1. Prioritisation: Both processes involve selecting and prioritising information based on relevance and importance. In humans, this prioritisation is influenced by factors such as emotional salience, attention, and repetition, while in AI, it is determined by the algorithm's design and training data.

  2. Strengthening connections: Just as reflection helps humans strengthen neural pathways and reinforce learning, summarising in AI can help to reinforce connections between related concepts and information. The act of summarising requires the AI to establish relationships between different pieces of information, thereby enhancing its understanding of the underlying concepts.

  3. Efficiency: Both reflection and summarising serve as mechanisms for making memory systems more efficient. By focusing on the most relevant and important aspects of information, both humans and AI can reduce the cognitive load and improve overall performance.

Retrieval Distortions: A Little Bit of Bias in All of Us

Imagine our memories as a game of historical telephone, where the message gets distorted as it's passed along. Humans tend to recall memories based on cognitive biases. AI systems, while not gossiping at a party, can also distort information when recalling it due to their own set of biases. So, whether you're a human or an AI, recalling memories can be a tricky business!

You know how we humans can be biassed when recalling memories, only remembering things that make us feel good or confirm our beliefs? Well, AI systems are no different. They can be biassed too, albeit for different reasons. AI's biases might come from energy efficiency, minimal token counts, algorithmic limitations, or even codec distortions due to data condensation techniques.

In both human and AI memory systems, biases influence the way information is retrieved and interpreted. Biases, like confirmation bias or selective memory, affect human recall, while AI biases might prioritise information that requires fewer computational resources or align better with training data.

Grouping Memories: A Fuzzy Affair
Creating summaries of summaries in AI is like your great-aunt's storytelling. Over time, she groups memories together, and the details become fuzzier with each retelling. Similarly, as AI condenses information further, the result is a less precise representation of the original data. Both human and AI summaries streamline memory by grouping and abstracting information, leading to a more manageable but less accurate picture.

This process is akin to the cognitive process of simplifying, generalising, and combining related memories into larger mental "chunks" in humans. In AI systems, the creation of summaries further condenses the information, leading to a potential loss of nuance and granularity.


In this memory adventure, we've discovered that human and AI memory systems share some surprising similarities and differences. Human memory and AI memory differ significantly in terms of their mechanisms and processes. While human memory relies on complex biological processes, AI memory is based on algorithms and data structures. Vector databases and summarising are two essential techniques for AI memory, enabling efficient storage, retrieval, and manipulation of information. As AI technology continues to evolve, understanding and refining these memory techniques will be crucial to unlocking the full potential of artificial intelligence systems.

The processes of summarising in AI and reflection in human memory share common features, such as prioritisation, strengthening connections, and improving efficiency. They are also strikingly similar in that they lead to a reduction in detail and precision. Biases also play a role in the recall process for both humans and AI, shaping the way information is retrieved and interpreted. Understanding these parallels can not only inform the development of more effective AI algorithms but also enrich our understanding of human memory and cognition.

Sidebar: AI memory
While this article focuses on LLM summarising techniques, there’s a lot more to AI memory. AI memory systems utilise various types of memory structures and techniques to store, manage, and process information. Some key types of memory used in AI include:

Explicit Memory
Explicit memory in AI systems refers to the storage of structured data, such as facts or rules, that can be accessed and used directly by the AI. This type of memory is similar to human declarative memory, which stores factual knowledge and can be consciously recalled.

Implicit Memory
Implicit memory in AI involves learning patterns or relationships in the data without explicit knowledge of the underlying structure. This is similar to human procedural memory, which involves learning skills and habits that are performed unconsciously. AI systems can develop implicit memory through unsupervised learning techniques, such as clustering, dimensionality reduction, or autoencoders.

Episodic Memory
Episodic memory in AI refers to the storage of specific events or experiences, allowing the AI system to recall past experiences to inform its decision-making process. This is analogous to human episodic memory, which involves the storage and retrieval of autobiographical events. AI systems can use techniques such as memory-augmented neural networks or reinforcement learning to develop episodic memory.

Semantic Memory
Semantic memory in AI is the storage of general knowledge about the world, concepts, and relationships between them. This is similar to human semantic memory, which deals with the meanings and understandings of words, objects, and ideas. AI systems can develop semantic memory through supervised learning, unsupervised learning, or knowledge graphs.

Short-term Memory
Short-term memory in AI refers to the temporary storage and manipulation of information required for immediate tasks or computations. This type of memory is similar to human working memory, which is responsible for maintaining and processing information for a short period. AI systems can use techniques such as recurrent neural networks (RNNs) or Long Short-Term Memory (LSTM) networks to implement short-term memory.

Long-term Memory
Long-term memory in AI involves the storage of information over extended periods. AI systems can use various data storage and retrieval techniques, such as databases or vector databases, to implement long-term memory. This type of memory is analogous to human long-term memory, which stores information over a more extended period.

Attention Mechanisms
Attention mechanisms in AI help focus the system's computational resources on the most relevant parts of the input data, similar to how humans selectively focus their attention. Attention mechanisms are particularly useful in deep learning models, such as transformers, to improve the efficiency and accuracy of the AI system by dynamically weighting the importance of different pieces of information during processing.

These different types of memory in AI systems allow them to learn, reason, and make decisions effectively by mimicking various aspects of human memory and cognitive processes.