Memory is any measurable change in a system that has a direct causal relationship to an experienced event. In other words, memory is any persistent change to a system (i.e., its properties, characteristics, or attributes) that is caused by an externally experienced event.

Memory (the caused change to the characteristics of a system) can be permanent or temporary. In order to be memory, the change must remain for some period after the causal event stops.



In computers

Memory is represented and implemented as a group of cells that are capable of holding a Boolean (logical) 1 or 0. The Boolean state of each bit is usually represented as a voltage, such as +5 volts to mean 1 (true), or 0 volts to represent 0 (false). Because these values can have no valid intermediate values they are binary, hence the memory cells represent binary bits. A 1 or zero bit can be written into an individually accessible cell, and the state of the specific cell can be polled at a later time.

Note that the thing that makes a computer memory valuable is not simply that it is capable of holding logical bits (which are in turn capable of encoding information), but that each bit is individually addressable. That is, there is a behavior associated with a group of memory that must also be implemented in order to make the memory usable. In essence, this behavior (repeatable addressing of individual bits within the collection) is just as important to, and therefor just as much a part of, the abstract concept: memory.



In biology

Memory in biological systems is also implemented through behavior. In the case of populations of individual cells, memory is achieved through attrition, which is the base mechanism in the process of adaptation. Attrition and adaptation also produce memory at the higher level, in groups of organisms, which act as individual cells within societies and civilizations.



In neurobiology

Neuron cells have processes which act, functionally, as individual cells do. That is, a neuron-cell has processes which can themselves, come and go, without the entire cell having to die or reproduce. For example, neurons form synapses, which are essentially information connections from other neurons. These connections can experience adaptation. They can completely die off, or new connections can form. In essence, the synapses themselves take the place of entire cells. This permits the cells --with all their complex and costly machinery-- to remain, while their synapses engage in the adaptive processes normally associated with cell-death, birth, and development.

Synapses can also model adaptation in a more continuous fashion. Not only can synapses form and die off (taking the place of cell birth and death respectively). They can also modulate the strength of their interconnection, based on adaptive pressures.

At the other extreme, memory and learning based on adaptation via cell-death is also seen in biological brains. Most (but not all) such learning occurs at prenatal or early neonatal stages of development, when entire populations of neurons are being born and then dieing off.



In Neural Networks

Memory is usually implemented via altering the strengths of synaptic connections between neurons. These connections are usually represented with computational-numeric values called weights, or connection-weights. They may also be represented in other ways, including, for example, the conduction or resistance of active or passive elements within an electronic circuit.



. . . . . . .
Resources

• Memristors (blog entry)