Scenario 1.

"What's the average rainfall in the Amazon Basin?"

Person: What?  Why do you ask?  I think it is a lot, maybe over 10 feet.

Computer: 150 inches.

Scenario 2:

"Did the bullies hit you?"

Person: No.

Computer: No records found.

"Did the bullies hit you here?"

Person: No.

Computer: No records found.

"No?  OK then how did you feel when the bullies hit you here?"

Person: I didn't like it.

Computer: No records found.

"So then what did you do?"

Person: I told them to stop.

Computer: No records found.

Human memory is imprecise and malleable.  It is fallible and susceptible to changing under leading questioning.  Even fully grown adults have shifting memory.  They may not recall what they saw, recall what they did not see, and  combine multiple memories into one (Ceci and Brook, 1998).  Computer memories in contrast are literally electronic file cabinets.  They store documents and records unchanged for as long as their underlying paper or magnetic media remain intact. 

The question here is: Do we wish to make human memory like the computer?  Or do we wish to make computer memory like the human?  And why would we want this?

Computer memories are electronic file cabinets.  Whatever we wish to store, it follows the engineer's coded instruction set to assign and write the file to a portion of its memory space.  More memory capacity means a larger file cabinet.  On shopping for computer memory / file cabinets, we may ask questions like, "How much can it store?" or "How long can it store it?"  Computer memory is well understood.  We created it.

Understanding human memory on the other hand is a work in progress.  But we must understand it better in order to answer whether we wish to make human memory into computer or vice versa.  The best place to start is with early, simple human memory.  The memories of young children.

Children from the earliest ages can readily understand that gist memory is more stable than verbatim memory (Brainerd, Reyna, Howe, and Kingma, 1990).  That is,  human memory is best at storing the main idea concepts rather than the actual video-like recording of an event.  To illustrate this more dramatically, Leichtman and Ceci (1995) examine the memory formation of preschoolers aged 3-6.   A teacher read them the fictional story of Sam Stone, a clumsy oaf who keeps knocking objects over and making a mess.  Later on, a guest visits the class.  His name is given as Sam Stone.  It was a typical nice, friendly visit.  That afternoon, the preschoolers find several objects have been knocked over and are covered in messy inks or are slightly torn.  The teacher asks, "Who did this?"

At first, the preschoolers did not know.  None claimed the guest, Sam Stone did it.  Until the repeated, leading questioning.  "Were you looking when Sam Stone went to the back and bumped into the teddy bear?"  Three quarters of the children claimed afterwards that Sam Stone had created the mess.  Half claimed they specifically saw Sam Stone do it.  One quarter even insisted strenuously that they were telling the truth and saw Sam Stone do it even when challenged.  When 100 external researchers and clinicians specialized in courtroom eye witness testimony with young children viewed the interviews, they were unable to determine which children were accurately telling the truth.  Conclusion: the children kept the gist memory and filled in the rest with a suitable narrative. 

Moving towards the very beginnings of human memory formation, Nadel and Zola-Morgan (1985) and Nelson (1995) explore the phenomenon of infantile amnesia.  Adults cannot recall any personal memories from 0-3 years of age.  Yet, children aged 0-3 definitely form memories about their moms, dads, siblings, room, crib, toys, walking, and diapering.  Conclusion: 0-3 year olds do learn yet either lose that learning or are unable to retrieve it later on.  Memories formed after about 4 years of age are permanent, however. 

Strauss and Cohen (1978) demonstrated that young children shown a series of images including features such as shape, color, size, and orientation (e.g. an Arrow, Red, Large, Pointing Up) retain the shape feature the longest.  This is followed by the color information, and least retained are the size and orientation.  Conclusion: On image memory, shape is the most salient feature.

Dempster (1981) explored random number sequence memorization by age.

This led to others (e.g. Pascual-Leone, 1989) to propose that memory capacity doubles between childhood and adulthood. 

The tentative, first look conclusion is that human memory is malleable and changing.  It is unreliable at early ages and wildly imprecise.  Young children, for example, can remember some things very well (e.g. image shape or the gist idea that Sam Stone is clumsy) and others not well at all (e.g. image size or the eye-witness details on Sam Stone). 

A second look yields more actionable frameworks.  Carr, Kurtz, Schneider, Turner, and Borkowski, (1989) explored how children of different ages memorized a list of terms:

couch, banana, dog, chair, apple, rat, table, cow, orange

Older children were able to remember far more accurately than younger children.  They were also able to better explain how they did it.  They organized them.  They grouped the list items by category:

Furniture (couch, chair, table)

Fruit (banana, apple, orange)

Animal (dog, rat, cow)

Similarly, Chi (1978) compared whether young chess prodigies were able to show the same level of skill across other domains and vs. ordinary adults.  The study asked both young prodigies and ordinary adults to memorize and reconstruct both chess positions and random numbers. 

The findings indicated that being expert at chess helped young chess prodigies better reconstruct chess positions, but provided no benefit for memory on ordinary random numbers.

Conclusion: background content knowledge helps form specialized framework that enables better memory.  In other words, we remember what we like to remember.  As a quick experiment, try to memorize this list:

sofa, platano, perro, silla, manzana, rata, mesa, vaca, naranja

Suddenly, the list is more difficult to remember since we may have difficulty categorizing it.  Unless we understood Spanish.  As with Socio-Cultural Learning Theory and natural language understanding, knowing a language is not just about mapping vocabulary or having the code phrases.  The language itself shapes our framework.  This framework is as important for us humans as physical framework is for a structure. 

Perhaps it is not simply a function of physical brain growth factors but more the scaffolding that draws our attention to form memories relevant to our anticipated needs.  This leads us to 3 basic points.

Point 1:

Memory is malleable and dynamic.  It is an active recording, not a passive one.  Our human memories are not the results of a staring array, but of attentionally focused senses.  We cannot do memory without a decision on what to record.  This is internalized social learning.  We remember based on what we earlier decided.  We earlier made the decision to look based on what we remembered.  This is necessarily a feedback loop, arguably the bane of modern computational systems.  This implies that our memory itself is a marker at least as informing as the fidelity of the memory.  It shows what we chose to remember, which shows what we chose to observe.

Point 2:

Memory is fit to a framework domain.  This is like formatting the computer file memory.  However, it does not help with other domains.  It may even - in the case of Artificial Intelligence scripted or computational tools - interfere with other domains.  There is a concept called crosstalk wherein having a system exposed to different but related domains can corrupt both data sets.  An earlier experiment at the University of Chicago working with Massachusetts Institute of Technology's ConceptNet trained for and completed an IQ test.  It scored as well as an average 4-year old.  Unfortunately, when drilling down into the results, it performed really well in passive recall, and atrociously poorly in inference and comprehension.  So if we want to look good with strong numerical results without deeper knowledge of what we humans do and want, then a computer purpose built for a particular task is fine.  This is a scaled up version of "teaching to the test."  But this is invariably useless to the business-side since business is not a pre-defined artificial test.  Hence the eternal mismatch between any company's business and technology departments.

Point 3:

Computers get written with a predefined framework. 

We humans form the framework on the fly.

Try this experiment: organize these articles by similarity.  There are no further instructions.

This sounds like an unsupervised machine learning clustering problem.  One can quote k-Means, hierarchical clustering, or statistical Expectation-Maximization.  But all these need more instructions.  How many clusters?  By what features?  For what goal?

Here is a secret: if you were able to organize the above list, you essentially decided what gets published.  Congratulations, you just did an editor's primary task.

Remembering a list is deciding how to organize these articles.  Therefore, memory is not passive high fidelity storage.  Memory in fact is not memory at all as we have come to know it within the common computational vocabulary.  It has become a misnomer.  Memory starts with deciding what to remember.  It continues with deciding how to decide what to remember.  We humans form frameworks to cluster on the fly because we decide what goes where.   We do this because we anticipate what we need for future narratives, stories, and entertainment.  We do this as children to make our caregivers happy.  E.g. Why do we remember the shape information best - at least in a Westernized, Anglophone culture?  Because in such a place, we caregivers care most about getting the shape correct. 

And we do this as adults to make our customers happy - to sell magazines, articles, and to close the deal.  In short, memory of the passive, storage, computational form is comparatively useless for business.  It is our ability to extract and add value by deciding what the customer wants that makes the framework for our malleable, imprecise human memory.  Computational memory is computational hardware centric.  Human memory is business centric.