STORAGE AND RETRIEVAL PROCESSES IN LONG-TERM

Psychological

Review

1969, Vol.

76, No. 2,

179-193

STORAGE

AND

RETRIEVAL PROCESSES

IN

LONG-TERM

MEMORY

1

R. M.

SHIFFRIN

2

AND

R. C.

ATKINSON

Stanford

University

A

theory

of

human memory

is

described

in

which

a

distinction

is

made

between three memory

stores:

the

sensory register,

the

short-term store,

and

the

long-term store. Primary emphasis

is

given

to the

processes

by

which

information

is

stored

in and

retrieved

from

the

long-term

store,

a

store which

is

considered

to be a

permanent repository

for

information. Forgetting

and

related phenomena

are

attributed

to a

failure

of the

retrieval process,

in

which

the

search through some memory area becomes less

efficient

as new

information

is

placed

in it.

Storage

and

retrieval

in the

long-term store

are

conceived

of as

parallel processes,

one

mirroring

the

other,

and

each

is

divided

into three stages

for

conceptual clarity.

The

memory trace

is

viewed

as an

ensemble

of

information

stored

in

some memory location,

the

location

of

storage determined largely

by the

components

of the

ensemble

itself.

The

ability

of the

system

to

cope with diverse phenomena

is

demonstrated

by a

consideration

of a

number

of

selected

experimental

paradigms.

Theories

of

human long-term memory

have given primary emphasis either

to the

organization

of

that

memory,

in

terms

of the

dimensions

of

storage

and the

associations

between

the

stored

information

(e.g., Cofer,

1965;

Deese, 1966; Mandler, 1968; Osgood,

1963),

or to the

characteristics

of

temporal

decay

from

that memory

as in the

interfer-

ence

theories (e.g., Keppel, 1968; Melton,

1963; Postman,

1961;

Underwood,

1957).

The

processes

by

which information

is

stored

in, and

retrieved

from,

long-term

memory

have been relatively neglected.

An

example

of

this type

of

process

is the

memory search

during retrieval; during

the

search,

a

suc-

cession

of

memory codes

is

examined,

each

examination followed

by a

decision process

in

which

the

search

is

either terminated

or

continued,

and in

which

information

recov-

ered

is

either accepted

as

that desired

and

1

Preparation

of

this article

was

supported

by a

grant

from

the

National Aeronautics

and

Space

Administration,

and is an

outgrowth

of

ideas

first

developed

in two

earlier reports (Atkinson

&

Shif-

frin,

1968b;

Shiffrin,

1968).

This

paper,

in

combi-

nation

with papers

by

Atkinson

and

Shiffrin

(1965,

1968a), represents

an

attempt

to

formulate

a

gen-

eral schema within which

to

analyze memory

and

learning.

2

Now at

Indiana

University.

Requests

for

re-

prints should

be

sent

to

Richard

C.

Atkinson,

Institute

for

Mathematical Studies

in the

Social

Sciences, Ventura

Hall,

Stanford University,

Stan-

ford,

California 94305.

output,

or

rejected.

It is the

intention

of

this paper

to

elaborate

the

memory input

and

output

processes.

As

will

be

indicated

later,

our

view

of

storage

and

retrieval eliminates

the

necessity

for

assuming decay

of

informa-

tion

from

long-term memory.

It

will

be

assumed that long-term memory

is

perma-

nent;

decrements

in

performance over time

are

ascribed

to an

increasingly

ineffective

search

of the

stored information.

We

begin

by

describing

the

overall concep-

tion

of the

memory system.

The

system

follows

that described

in

Atkinson

and

Shif-

frin

(1965,

1968a),

and is

similar

to

those

proposed

by

'Feigenbaum

(1966)

and

Nor-

man

(1968).

The

major components

of

the

system

are

diagrammed

in

Figure

1:

the

sensory register,

the

short-term store (STS)

and

the

long-term store

(LTS).

The

solid

arrows

in the

diagram represent directions

in

which information

is

transferred

from

one

part

of the

system

to

another. Note that

transfer

is not

meant

to

imply

the

removal

of

information

from

one

store

and the

placing

of

it in the

next;

rather, transfer

is an

opera-

tion

in

which information

in one

store

is

"copied" into

the

next without

affecting

its

status

in the

original

store.

It

should

be

emphasized

that

our

hypotheses

about

the

various

memory stores

do not

require

any

assumptions

regarding

the

physiological

locus

of

these

stores;

the

system

is

equally con-

179

180

R. M.

SHIFFRIN

AND R. C.

ATKINSON

Stimulus

analyzer

programs

Activate

rehearsal

mechanism

Modify

information

flow

from

SR I

Code

and

transfer

information

fror

Initiate

or

modify

search

of LTS

Initiate

response

generator

.

STS

li

FIG.

1. A flow

chart

of the

memory

system.

(Solid

lines

indicate

paths

of

information

trans-

fer.

Dashed lines indicate connections which permit comparison

of

information arrays residing

in

different

parts

of the

system; they also indicate paths along which control signals

may be

sent

which activate information transfer, rehearsal mechanisms, etc.)

sistent with

the

view that

the

stores

are

separate physiological

structures

as

with

the

view

that

the

short-term store

is

simply

a

temporary activation

of

information perma-

nently

stored

in the

long-term store.

The

control processes listed

in

Figure

1 are a

sample

of

those which

the

subject

(S) can

call

into play

at his

discretion, depending upon

such

factors

as the

task

and the

instructions.

Control processes govern informational

flow,

rehearsal,

memory search, output

of re-

sponses,

and so

forth.

The

sensory register

is a

very short-lived

memory

store which temporarily holds

in-

coming

sensory

information while

it is

being

initially

processed

and

transferred

to the

short-term store.

In the

visual modality,

for

example,

information

will

decay

from

the

sensory register

in a

period

of

several hun-

dred milliseconds

(Sperling,

1960).

Infor-

mation

in the

short-term store,

if not

attended

to by S,

will

decay

and be

lost

in a

period

of

about

30

seconds

or

less,

but

control proc-

esses such

as

rehearsal

can

maintain informa-

tion

in STS for as

long

as S

desires (the

buffer

process

in

Figure

1 is one

highly

organized rehearsal scheme).

While

infor-

mation resides

in

STS, portions

of it are

transferred

to

LTS.

The

long-term

store

is

assumed

to be a

permanent repository

of

information;

we

realize that factors such

as

traumatic

brain damage, lesions,

and de-

terioration with extreme

age

must lead

to

memory loss,

but

such

effects

should

be

negligible

in the

types

of

experiments con-

sidered

in

this paper.

Thus

it is

hypothe-

sized

that information, once stored

in

LTS,

is

never thereafter destroyed

or

eliminated.

Nevertheless,

the

ability

to

retrieve informa-

tion

from

LTS

varies considerably with time

and

interfering

material.

The

short-term

store

serves

a

number

of

useful

functions.

On the one

hand

it de-

couples

the

memory system

from

the

external

environment

and

relieves

the

system

from

the

responsibility

of

moment-to-moment

at-

tention

to

environmental changes.

On the

other hand,

STS

provides

a

working memory

in

which manipulations

of

information

may

take place

on a

temporary basis. Because

STS is a

memory store

in

which information

can be

maintained

if

desired,

it is

often

used

as the

primary memory device

in

certain

types

of

tasks;

in

these tasks

the

information

PROCESSES

IN

LONG-TERM

MEMORY

181

presented

for

retention

is

maintained

in STS

until

the

moment

of

test

and

then emitted.

Tasks

in

which

STS is

utilized

for

this pur-

pose,

and the

mechanisms

and

control proc-

esses that

may

come into play, have been

examined extensively

in

Atkinson

and

Shif-

frin

(1968a).

In

this report

we are

pri-

marily

interested

in STS as a

temporary

store

in

which information

is

manipulated

for

the

purposes

of

storage

and

retrieval

from

LTS, rather than

as a

store

in

which

infor-

mation

is

maintained until

test.

In the re-

mainder

of

this paper, discussion

is

limited

to

that component

of

memory performance

which involves

LTS

retrieval,

and the

com-

ponents

arising

from

STS and the

sensory

register

will

not be

considered.

LONG-TERM

STORE

In

describing

the

structure

of

LTS,

an

analogy with computer memories

is

helpful.

The

usual computer memory

is

"location

addressable";

if the

system

is

given

a

certain

location

it

will

return

with

the

contents

of

that location. When given

the

contents

of

a

word

(a

"word"

refers

to a

single com-

puter memory

location),

such

a

system must

be

programmed

to

examine each location

in

turn

in

order

to find the

possible locations

of

these contents

in the

memory.

It

seems

untenable that

an

exhaustive serial search

is

made

of all of LTS

whenever retrieval

is

desired.

An

alternative type

of

memory

may be

termed

"content-addressable";

if the

system

is

given

the

contents

of a

word

it

will

return with

the

locations

in

memory

containing those contents.

One way in

which

such

a

memory

may be

constructed utilizes

a

parallel search through

all

memory loca-

tions;

the

system then returns with

the

locations

of all

matches.

If

this view

is

adopted, however,

an

additional process

is

needed

to

select

the

desired location from

among

the

many returned

by the

parallel

search.

Thus,

if we

feed

the

system

the

word

"red,"

it

would

not be

useful

for the

system

to

return with

all

references

or

locations

of

"red";

there

are

simply

too

many

and the

original

retrieval

problem would

not be

sig-

nificantly

reduced

in

scale. There

is,

how-

ever,

an

alternative method

for

forming

a

content-addressable

memory;

in

this

method,

the

contents

to be

located themselves contain

the

information necessary

to

specify

the

stor-

age

location(s).

This

can

occur

if the in-

formation

is

originally stored

in

locations

specified

by

some master plan dependent

upon

the

contents

of the

information. Such

a

system

will

be

termed "self-addressing."

A

self-addressing memory

may be

compared

with

a

library shelving system which

is

based

upon

the

contents

of the

books.

For

example,

a

book

on

"caulking methods used

for

12th

century

Egyptian

rivercraft"

will

be

placed

in

a

specific

library location

(in the

Egyptian

room,

etc.).

If a

user desires

this

book,

it

may

be

located

by

following

the

same

shelv-

ing

plan used

to

store

it in the first

place.

We

propose that

LTS is to a

large degree

just such

a

self-addressable

memory.

An

ensemble

of

information

presented

to the

memory

system

will

define

a

number

of

memory

areas

in

which that information

is

likely

to be

stored;

the

memory search will

therefore

have certain natural starting points.

The

system

is

assumed

to be

only partially

self-addressing

in

that

the

degree

to

which

the

storage locations

are

specified

will vary

from

one

ensemble

to the

ment

to the

next,

in

much

the way as

pro-

posed

in

stimulus sampling theory

(Estes,

1959).

Thus

it may be

necessary

to

embark

upon

a

memory search within

the

specified

locations,

a

search which

may

proceed seri-

ally

from

one

location

to the

next.

This

conception

of LTS

leads

to a

number

of

pre-

dictions.

For

example,

a

recognition test

of

memory

will

not

proceed

via

exhaustive scan-

ning

of all

stored codes,

nor

will

a

recognition

test eliminate

in all

cases

the

necessity

for an

LTS

search.

If

information

is

presented

and

S

must indicate whether

this

information

has

been

presented previously, then

the

likely

storage

location

(s) is

queried.

To the de-

gree

that

the

information

has

highly

salient

characteristics which precisely

identify

the

storage location,

the

extent

of the LTS

search

will

be

reduced.

Thus,

for

items with highly

salient

characteristics,

6"

should

be

able

to

identify

quickly

and

accurately whether

the

item

was

presented previously,

and the

iden-

tification

might

not

require

a

memory search

which

interrogates more than

a

single storage

location.

The

less

well-specified

the

storage

182

R. M.

SHIFFRIN

AND R. C.

ATKINSON

location,

the

greater

the

memory search

needed

to

make

an

accurate recognition

re-

sponse.

This

view

is in

many respects simi-

lar to

that proposed

by

Martin

(1967),

in

which

he

suggests

that

certain stimuli tend

to

give

rise

to

identical encoding responses

from

one

instance

to the

next,

that

is, on

both

study

and

test trials.

From

the

present

viewpoint

"identical encoding responses"

is

taken

to be

equivalent

to

identical storage

locations.

Although

it is

assumed

that

information

ensembles

are

sorted into

LTS

locations

ac-

cording

to a

master plan,

it is

beyond

the

scope

of

this paper

to

attempt

to

outline this

organizational

structure;

other workers have

been

dealing with this problem (e.g.,

Han-

dler,

1968;

Pollio,

1966).

Undoubtedly this

organization

is

highly complex,

but the

form

of

the

organization will

not be

crucial

to our

discussions

of

input

and

output mechanisms.

There

is one

dimension

of

organization

that

must

be

mentioned here,

however;

this

is the

"temporal"

dimension.

We are

assuming

that this dimension

is

like

the

other organi-

zational

dimensions

in

that information

may

be

stored along

it, and

that retrieval

may be

based upon

it. In the

remaining sections

of

this paper,

the

temporal dimension

will

often

be

singled

out for

special mention,

not be-

cause

it

differs

in

substance

from

other

organizational dimensions,

but

because

it is

virtually

impossible

to

eliminate

it as a

sys-

tematic variable

in

most memory

experi-

ments.

The

term

"location"

is

used

in

relation

to

the

organizational

schema;

an LTS

location

is

defined

by the

place

in the

organizational

structure occupied

by an

information

en-

semble.

The

location will

be

defined

in

terms

of

the

modality

of the

information

(e.g.,

visual versus

auditory),

the

level

of

analysis,

(e.g.,

spelling versus syntactic

structure),

and all

other

dimensions

of

organization

that

may

be

relevant.

Two

locations will

be

said

to be

"close"

if the

information

in

them tends

to be

retrieved together.

In

particular,

we

shall

refer

to a

code,

or an

image,

as an

ensemble

of

information

that

is

closely related

and

very

likely

to be

retrieved together.

We

do not

wish

to

imply that there

is

some uni-

tary atom

of

storage called

a

code

or an

image.

The

information making

up a

code

in

one

task

may be

considered

to be

several

codes

in a

different

task.

'For

example,

an

entire sentence

may be

considered

a

code

if

we

are

comparing

the

meaning

of

that sen-

tence

with

others;

however,

the

same sen-

tence

might

be

considered

to be

made

up of

a

series

of

codes

if we are

comparing

it

with

sentences

of the

same meaning

but

different

grammatical

form.

Nevertheless,

for

most

tasks

the

concept

of a

code

or

image

as

rep-

resenting

a

cohesive array

of

information

in

a

single storage location proves

useful.

This

view

of the

memory trace

is

consistent with

the

analysis given

by

Bower

(1967).

In his

model,

a

trace,

or

image,

is

represented

by a

vector

of

attributes which serves

to

identify

both

the

information contained

in the

trace

and the

location

of the

trace

in

LTS.

The

values

of the

attributes

in the

vector serve

to

indicate

the

position

of the

trace

on the

vari-

ous

organizational

dimensions;

for

example,

one

of the

positions

in the

vector might indi-

cate whether

the

trace

is

auditory

or

visual.

It

should

be

apparent that this quantitative

view

is

compatible with

our

description

of

an

image

or

code,

and

also with

the

descrip-

tion

of LTS as a

self-addressable store.

STORAGE

AND

RETRIEVAL

Since

LTS is

self-addressing, storage

and

retrieval have many features

in

common,

one

process

mirroring

the

other.

Storage

is

assumed

to

consist

of

three primary mecha-

nisms:

transfer,

placement,

and

image-pro-

duction.

The

transfer mechanism includes

those control processes

by

which

5"

decides

what

to

store, when

to

store,

and how to

store

information

in

LTS.

The

placement mecha-

nism

determines

the

locations

in

which

the

ensemble

of

information

under consideration

will

be

stored.

To a

large degree,

the

com-

ponents

of the

ensemble itself

will

determine

the

location

of

storage.

That

is, in the

action

of

encoding

the

desired information

for

stor-

age,

5"

may

supplement

the

information

cur-

rently

in STS

with pertinent

information

retrieved

from

LTS;

the

resultant ensemble

in

STS

determines

the

storage location.

The

image-production

mechanism

determines

what proportion

of the

current ensemble

of

information

in STS

will

be

placed

in the

PROCESSES

IN

LONG-TERM

MEMORY

183

designated

LTS

location(s).

The

propor-

tion stored should

be a

function

of the

dura-

tion

of the

period that

the

ensemble

is

main-

tained

in

STS. Retrieval, like storage,

is

assumed

to

consist

of

three primary mecha-

nisms

:

search, recovery,

and

response gen-

eration.

The

search process

is a

recursive

loop

in

which locations

or

images

are

suc-

cessively selected

for

examination.

As

each

image

is

examined,

the

recovery process

de-

termines

how

much

information

will

be

recovered

from

the

image

and

placed

in

STS.

The

response generation process then exam-

ines

the

recovered information

and

decides

whether

to

continue

the

search

or

terminate

and

emit

a

response.

If the

search does

not

terminate,

the

selection

of the

or

image

for

examination

may

depend upon

information

already uncovered during

the

search.

Although storage

and

retrieval

are

treated

separately

in

this paper,

we do not

wish

to

imply that these processes

are

separated

in

time,

one

following

the

other. Rather, long-

term storage

is

continually occurring

for the

information

residing

in

short-term store.

In

addition, retrieval

is

continually occurring

during

storage

attempts

by S; for

example,

5*

may try to

store

a

paired-associate

by

searching

LTS for

prominent associations

to

the

stimulus, associations which could then

be

used

as

mediators.

The

remaining portions

of

this report will

be

directed toward

a

delineation

of

these

in-

put

and

output processes.

It is

beyond

the

scope

of

this paper

to do

more than

briefly

describe

the

major

mechanisms

of the

theory.

Nevertheless,

an

attempt will

be

made

to

indicate

how

these processes work, what evi-

dence supports

our

assumptions, what pre-

dictions

may be

derived,

and how the

theory

may be

applied

to a

number

of

selected

tasks.

Storage Processes

Transfer.

The

decisions involving what

to

store, when

to

store,

and how to

store infor-

mation

are

under

a

high degree

of

control

by

S, and

therefore

can

result

in

striking per-

formance

changes

from

one

task

to

another.

Although

we

shall

be

discussing experimental

situations

in

which

storage takes place largely

during designated study periods,

we

should

note that

storage

in

general occurs whenever

information

is

cycled through

the

short-term

store;

for

example,

in

reflective thinking,

in

daydreaming,

and in

moment-to-moment

consideration

of the

day's

happenings.

The

decision concerning when

to

store

information

is

especially important

in

situa-

tions where

a

large amount

of

information

is

being input rapidly

to

STS. Such

a

situa-

tion

taxes

the

capacity

of the

system

and

forces

5"

to

select

a

subset

of the

presented

information

for

special attention

and

coding.

There

are a

number

of

factors

that

determine

the

information

so

selected. Important

or

easily

stored

information

is

likely

to be

given

preference.

For

example,

Harley

(1965)

has

shown

in a

mixed-list

design

that

paired-

associate items given high monetary

payoffs

are

selectively attended

in

preference

to

items

given

low

payoffs.

Information selection will

also

be

governed

by the

degree

to

which

the

incoming

material

is

already known.

For

example, there

is

considerable evidence that

5s

tend

to

store more

information

on a

given

item

if

that

item's

presentation

is

fol-

lowed

by

other items that

are

well-known

as

opposed

to

being

followed

by

items that

are

not

known (Atkinson,

Brelsford,

&

Shiffrin,

1967;

Thompson,

1967).

Finally, note that

the

decision when

to

store

may be

determined

by

strategies

associated

with

the

list

or

task

as a

whole, rather than with

the

individual

item.

For

example,

in an

experiment

em-

ploying

a

paired-associate list with only

the

two

responses

A and B, S

might decide

to

store only

information

about those stimuli

paired with Response

A, and

always guess

Response

B if the

answer

is not

known

at

test.

The

decisions concerning

how to

store

in-

formation

will

also

affect

performance:

stor-

age via

visual images

may be

more

effective

than auditory storage

(Schnorr

&

Atkinson,

1969);

overt

and

covert rehearsal methods

may

result

in

very

different

effects

(Brels-

ford

&

Atkinson,

1968);

and

mediating ver-

sus

nonmediating instructions

may

give rise

to

considerable performance

differences

(Runquist

&

Farley,

1964).

On a

somewhat

different

level,

5"

may

engage

in

organiza-

tional

storage strategies such that

different

items

are

stored

in

locations determined

by

184

R. M.

SHIFFRIN

AND R. C.

ATKINSON

some

fixed

organizational framework

(Tul-

ving,

1962).

What

information

is

stored, given

the

presentation

of a

particular information

en-

semble,

will also

be

highly dependent upon

the

control

processes

utilized

by S. In

some

cases,

as

Underwood

(1963)

points

out in

making

a

distinction between

the

nominal

and

functional

stimulus,

not all the

informa-

tion

contained

in the

presented item

is

neces-

sary

for

correct responding (e.g.,

if the

stimuli

are all

nonsense syllables

differing

only

in the first

letter, then only

the first

letters

need

be

stored).

In

these

cases

only

the

relevant characteristics

of the

input need

be

stored.

In

most cases

S

will select

relevant characteristics

of the

presented

in-

formation

and add to

this additional coding

information

from

LTS;

for

example,

a

paired-associate

plus

a

mediator might

be

stored

in

many cases.

One

type

of

informa-

tion

which

S

often

attempts

to

store

is in-

formation

indicating that

a

particular

re-

sponse (usually

one

just given

in

error)

is

not

correct;

Millward

(1964)

has

examined

evidence

for

this process

in

simple paired-

associate

tasks.

This

process

of

tagging

responses

as

incorrect

is

particularly impor-

tant

in

studies

of

negative transfer;

the

higher

the

probability that

the first

response

assigned

to a

stimulus

will

be

tagged

as in-

correct (when

the

response

changes),

the

less will

be the

proactive interference

effect

observed

in the

data

(Shiffrin,

1968).

The

examples

of

transfer mechanisms

given

above

are by no

means exhaustive,

but

they

serve

to

indicate

the

pervasiveness

and

importance

of

these processes. Relatively

simple

decisions

to

select

one

transfer scheme

rather than another

can and do

lead

to

large

performance

effects.

These

facts

emphasize

the

need

for

carefully

establishing

the

trans-

fer

mechanisms used before extending analy-

sis to

other aspects

of the

data.

Placement.

The

location

in

which

an

image

will

be

stored

is

determined

by the

contents

of

that

image;

5"

therefore controls

the

storage location

by

manipulating

the

information

complex

in

STS.

For any

given

ensemble

of

information, however, there

is

a

certain amount

of

randomness

in

placement.

For

this

reason,

a

search

for an

image

may

have

to be

undertaken

at

test

even

if the

entire

information ensemble originally

pres-

ent

in STS

during study

is

presented

for

consideration.

Information

to be

remem-

bered

may be

stored

in

images

in

more than

one

location;

for

example,

a

paired-associate

may be

encoded

by the use of two

entirely

different

mnemonics. This notion

has

been

given

quantitative

form

in

"multiple-copy"

models

for the

memory trace (Atkinson

&

Shiffrin,

1965).

The

primary mechanism determining

stor-

age

location

is in all

cases

an

organizational

framework.

The

self-addressing character-

istic

of

input information depends upon

the

prior, already established,

LTS

organization;

each

item

is

sent

to

locations that depend

upon

the

preexisting organizational

frame-

work. However,

in

many

cases

(especially

situations

involving free-recall learning

and

paired-associate learning

in

which

the

same

list

of

items

is

presented over

a

series

of

trials)

it is an

effective

technique

for

5

1

to

form

new

organizational structures.

In

free-

recall learning,

for

example, output

on

suc-

cessive

trials becomes increasingly organized

and

similar

to

output

on the

preceding trial

(Gofer,

1965;

Tulving,

1962);

it may be

inferred

that

a

consistent

new

organization

has

been imposed upon

a set of

words which

were

disorganized

at the

start

of

learning.

This

point regarding growth

of

organization

is

especially important with regard

to

list-

structured paired-associate tasks,

that

is, a

task

in

which

a

list

of

paired-associates

is

presented over

and

over

on

successive trials

(possibly

in a new

random order

from

trial

to

trial).

In

such

a

task, organizational

effects

between items will tend

to

occur over trials,

and

care must

be

taken

in

inferring that

effects

seen

by

averaging data over

all

items

in

the

list will also apply

to the

individual

items.

For

example, there

is

evidence indi-

cating

that interference

effects

found

by

averaging over items

in a

list

do not

apply

to the

individual items making

up the

list

(DaPolito, 1966; Greeno,

1967).

For

this

reason

we

shall

often

consider

in the

rest

of

this

paper

results

from

what

is

termed

a

"continuous"

paired-associate task (Bjork,

1966;

Brelsford,

Shiffrin,

&

Atkinson,

1968;

Rumelhart,

1967).

This

type

of

task

em-

PROCESSES

IN

LONG-TERM MEMORY

185

ploys

a

long series

of

presentations, each

involving

first a

test

and

then

a

study

on a

paired-associate item.

The

character

of the

experiment

is

essentially

homogeneous over

time

because

new

items

are

continually being

introduced

and old

items

deleted;

a new

item

may

appear

at any

point

in the

sequence,

receive

a fixed

number

of

presentations dis-

tributed over

a

subset

of

trials,

and

then

be

dropped

and

replaced

by yet

another item.

In

this

way the

list-structure feature

of the

typical paired-associate experiment

is

elimi-

nated,

and it is

extremely

difficult

for

-5"

to

develop

special schemes

for

interitem organi-

zation.

In

most cases,

5"'s

placement strategy

in-

volves choosing

one of

many preexisting

or-

ganizational dimensions

for

storage.

In

these cases,

an

organizational clue contained

in

the

experimental design

may

prove

useful.

In

categorized free-recall,

for

example,

5"

will

be

induced

to

store

(and retrieve)

in the

given categories

(Bousfield

&

Cohen,

1956;

Cofer,

1965;

Cohen,

1963).

Thus

the

loca-

tion

in

which

the

word division will

be

stored

will

be

quite

different

if

preceded

by

multi-

plication, addition,

and

subtraction than

if

preceded

by

platoon,

regiment,

and

battalion,

In the

interests

of

effective memory,

the

most important requirement

of the

placement

process

is

that

it

results

in a

storage location

which

will later

be

searched during test.

If

organizational

schemes

are

used

for

place-

ment,

the

schemes themselves

will

have

to be

stored,

recovered,

and

utilized

in

order

for

retrieval

to be

optimal.

Image-production. When

an

ensemble

of

information

is

present

in

STS, some portion

of

it

will

be

stored

in a

designated location

in

LTS as a

permanent image.

The

propor-

tion

of

information that

is

stored will

be a

function

of the

duration

of

time

that

the

ensemble stays

in

STS,

or of the

number

of

times

that

ensemble

is

cycled through STS.

It

would

be

most natural

to

look

for

evidence

of

this mechanism

in

experiments varying

study

time

for

particular items. However,

improved

performance with longer study

times

can be

attributed

to a

higher chance

of

finding a

good mnemonic. Better evidence

is

found

in

experiments

in

which

5"s

are in-

duced

to

utilize rote rehearsal rather than

coding,

and in

which longer periods

of

rote

verbal rehearsal lead

to

improved perform-

ance

(Brelsford

&

Atkinson, 1968; Hellyer,

1962).

To

conclude

the

discussion

of

storage,

we

consider

the

content

of the

image:

the

range

and

form

of the

stored information.

A

single image

may

contain

a

wide variety

of

information,

including characteristics

of the

item presented

for

study (its sound, meaning,

color,

size,

shape, position, etc.)

and

char-

acteristics added

by S

(such

as

codes, mne-

monics,

mediators, images, associations,

etc.).

In

addition,

an

image usually

will

contain

links

to

other images

(other

information

which

was in the

short-term

store

at the

same

time)

;

these links

can be

regarded

as a set

of

directions

to the

locations

of

ages

in

LTS.

Retrieval

Processes

The

retrieval mechanism forms

the

crux

of

the

present theory, since

it

enables

a

per-

manent

long-term store

to

exhibit

the

char-

acteristics

of a

failing memory.

The

basic

mechanism

by

which memory loss occurs

involves

a

partially random search through

an

increasingly large

set of

images

in

some

local

set of

memory

locations.

This

local

area

may be

defined

by one or

more dimen-

sions

in the

organizational structure

of

LTS;

as the

number

of

images

in the

local area

increases,

the

search

for the

desired informa-

tion

will

become increasingly

ineffective.

For

example,

consider

two

areas,

one

con-

sisting

of

three images,

the

other consisting

of

10

images,

and

both containing

the

desired

code.

A

random search through

the

smaller

set

will result

in

successful retrieval

in a

shorter period

of

time than

a

search through

the

larger

set.

The

retrieval process begins with

the

pre-

sentation

of an

information complex which

places constraints

on the

response desired

and

also provides

a

number

of

clues,

or

delimiting

information, concerning

that

de-

sired

output (i.e.,

a

stimulus).

On the

basis

of

this presented information,

or as the

result

of

an

external search strategy,

5"

is led to

look

in

some local memory

area

and

select

(possibly randomly)

an

image

for

examina-

tion.

The

process

by

which information

is

186

R. M.

SHIFFRIN

AND R. C.

ATKINSON

recovered

from

this image

is

called

recovery.

The

recovered information will

be

placed

in

the

short-term

store

which

may

also contain

other information such

as the

search

strategy

being employed, salient information recov-

ered previously

in the

search,

the LTS

loca-

tions

that

have been examined already,

and

some

of the

links

to

other images

that

have

been

examined already,

and

some

of the

links

to

other images that

have

been noted

in the

search

but not yet

examined.

The

short-

term

store

thus

acts

as a

"window"

upon

LTS, allowing

5"

to

deal sequentially with

a

manageable amount

of

information.

The

cur-

rent

contents

of STS are now

examined

and

various decisions made concerning whether

the

desired response

has

been

found,

whether

to

emit

it,

whether

to

terminate

the

search

unsuccessfully,

or

whether

to

continue

the

search.

These

decisions

and the

generation

of

the

response

are

called

the

response-gen-

eration

process.

If a

decision

is

made

to

continue

the

search, then

a new

location

is

selected either randomly,

on the

basis

of

information

just recovered,

or in

accord with

an

overriding external search strategy.

The

process which continues cycling

in

this man-

ner

until termination

is

called

the

search

process.

Search.

Each

cycle

in the

search recursion

begins

with

a

mechanism which locates

the

next image

for

examination.

This

mecha-

nism

may be

separated into directed

and

ran-

dom

components.

The

directed component

includes

strategies

controlled

by S and de-

pends upon

the

input information

and the

self-addressing

nature

of the

system.

As a

result

of the

directed component,

a

number

of

locations

in

some area

of

memory

are

marked

for

examination.

The

locations

and

images

so

marked will

be

referred

to as the

exami-

nation subset,

and the

directed component

of

the

search

may be

characterized

by the

proba-

bility

that

the

sought-after code

is in the

examination subset.

The

directed component

of the

search

can

be

either under

a

high degree

of

5

1

control

or

relatively automatic.

For

example,

the

control

factor

is

high

in a

situation where

5"

engages

in a first-letter

alphabetic search,

or

where

5"

attempts

to

remember lunch

of

2

days previously

by

reconstructing

the

events

of

that day.

In

cases like this,

the

search assumes many

of the

aspects

of

prob-

lem

solving.

At the

other

extreme

is the

almost completely automatic direction

of

search which occurs, say,

in the

attempted

recognition

of a

word

as

having been pre-

sented earlier

in the

session.

In

such

a

case,

the

word

itself

serves

to

direct

the

search

via the

self-addressing

feature

of the

system.

In its

most general

form

the

search process

can

be

viewed

as a

series

of

stages

in

which

searches

are

made successively

in

different

examination subsets.

That

is,

after

choosing

codes randomly

for a

time

from

one

subset,

information

recovered during

the

search

may

cause

5"

to

change

the

memory area currently

being examined

for one

quite

different.

This

type

of

search

could occur,

for

example,

in

a

categorized

free-recall

task,

in

which

the

various categories

are

searched successively

(Cohen,

1963).

In

most applications, how-

ever,

a

restricted search model should

be

ade-

quate,

the

restriction allowing only

a

single

examination subset

to be

searched during

any

retrieval period.

This

model should

be

particularly applicable

to

tasks

in

which only

a

short response period

is

allowed.

The

random component

of the

search

specifies

the

selection

of

codes

in the

exami-

nation subset

from

one

cycle

of the

search

to

the

next.

It is

assumed that

the

likelihood

of

any

particular code being selected

at

some

point

will

depend upon

the

amount

of

infor-

mation contained

in it, as

well

as the

total

number

of

codes

and

their temporal ordering

in

the

examination subset.

It

will

be

seen

that

the

extent

to

which

the

search depends

upon

the

temporal ordering

of the

items

is

an

important variable which, among other

things,

determines interference

effects.

The

division

of

search into directed

and

random

components leads

to

somewhat dif-

ferent

memory models, depending upon

which

component

is

selected

for

elaboration.

Feigenbaum

(1966)

and

Hintzman

(1968),

for

example, have elaborated upon

the di-

rected component

in

computer simulation

models.

In

these

formulations

there

is a

mechanism

termed

a

"discrimination net,"

which

enables

S to

sort through

the

organi-

zational

structure

of LTS to

reach

the

local

memory

area where desired information

is

PROCESSES

IN

LONG-TERM

MEMORY

187

stored.

On the

other hand, there

are

models

which emphasize

the

probabilistic

search

through

an

examination subset

of

items

in

some memory area (Atkinson

&

Shiffrin,

1965).

It is

upon this latter type

of

model

that attention will

be

focused

in

this paper.

Recovery.

Once

an

image

has

been

lo-

cated,

it is

appropriate

to ask

what informa-

tion contained

in the

image will

be

entered

into

the

short-term

store.

This

process

is

called

recovery.

The

amount

of

information

recovered

from

an

image

is

assumed

to be

probabilistic, depending upon

the

current

noise level

in the

system

and the

amount

of

information

in the

image.

In

particular,

the

amount

of

information

recovered should

be

an

increasing

function

of the

amount

of in-

formation

in the

image.

Response generation. Having recovered

information

from

LTS,

5"

is

faced

with deci-

sions

as to

whether

to

terminate

the

search

and

respond

or to

continue

to

search. These

decisions must

first of all

depend upon

the

consistency

of the

recovered information with

that

indicated

by the

test information.

In-

consistent

information

can be

ruled

out at

once.

If

information

consistent with

the

test

stimulus

is

found,

and a

recognition response

is

desired, then

the

response will

be

given

when

temporal

or

contextual information

is

recovered

indicating that

the

image

was

stored recently.

In a

paired-associate

test,

three

types

of

information

may be

recovered:

that

associated with

the

stimulus,

that

asso-

ciated with

the

response,

or

associative

in-

formation

linking

the

two. Recovery

of

stimulus

or

response

information will

serve

to

lead

to

recognition

of

either;

if

both

are

recovered

in

nearby locations

in

memory,

then

the

response

may be

emitted.

On the

other hand,

a

response might

be

output fol-

lowing recovery

of

associative information

linking

the

response

to the

stimulus. Since

the

ability

to

output

a

response depends upon

the

amount

of

response

information

recov-

ered,

the

process

may

often

be

represented

by

a

decision-theoretic model

in

which

S is at-

tempting

to filter

information through

a

noisy

background

(Bernbach,

1967;

Kintsch,

1967;

Wickelgren

&

Norman,

1966).

In

situations

where

a

long time period exists

for

respond-

ing,

a

likely response

may be

recovered

but

not

emitted;

in

these cases

the

search will

be

continued

in the

hope

of finding a

better

response.

An

extended

search

of

this kind

leads

naturally

to

predictions that

5"

will

be

able

to

rank responses

in the

order

of

their

probability

of

being correct, with responses

ranked after

the first

being

correct

at an

above-chance level (Binford

&

Gettys,

1965).

A

decision

can be

made

to

terminate

the

search

unsuccessfully

if

response

time

has

run

out,

or if

response time

is

expected

to

run out and

.S"

wishes

to

make

a

guess

before

it

does,

or if

5

decides that further search

would

not be

useful.

Termination schemes

of

the

latter type

are

quite varied.

One

simple

rule would terminate

the

search when

all

images

in the

examination subset have

been

interrogated

unsuccessfully.

Other

schemes would

end the

search when some

fixed

time

limit expired,

or

some

fixed

num-

ber

of

items examined.

These

schemes

are

described

in

more detail

in

Atkinson

and

Shiffrin

(1965).

The

extensive decision structure that

has

been

outlined

may

make

it

seem unlikely that

time

is

available

during

the

search

for

exami-

nation

of a

very large number

of

images.

This

is

indeed

the

point

of

view

we

adopt:

in

most cases

it is

assumed

that

only

a few

images will

be

examined

in the

search before

termination.

For

example,

in a

continuous

paired-associate task analyzed

by

Shiffrin

(1968),

the

estimated number

of

codes exam-

ined prior

to

termination

was

from about

one to five.

APPLICATIONS

OF THE

SYSTEM

Forgetting

Decrements

in

performance occur

in the

system

as a

result

of the

input

of

additional

information

to

LTS. These decrements

re-

sult

from

three related mechanisms.

First

is a

mechanical

effect;

information

sufficient

to

respond

correctly

at one

point

in

time

may

prove inadequate

after

additional information

has

been added.

For

example,

a

paired-

associate GAX-4

may be

stored

as

G**-4,

and

this code

will

be

sufficient

for

correct

responding

(if

recovered) when

GAX is

tested. Suppose, however, that

GEK-3

is

now

presented

and

stored

as

G**-3.

When

188

R. M.

SHIFFRIN

AND R. C.

ATKINSON

either

of

these stimuli

is

tested later, both

codes

may be

retrieved

from

LTS and

there-

fore

6"

will

have

to

guess whether

the

correct

response

is 3 or 4.

The

second cause

of

forgetting arises

from

a

breakdown

in the

directed component

of

the

search mechanism.

That

is,

correct

re-

trieval

requires that

the

same memory

area

be

searched

at

test

as was

used

for

storage

during

study.

This

may not

occur, however,

if

only

a

portion

of the

input

information

is

used

to

direct storage during study,

for a

different

portion might

be

utilized

to

locate

the

storage area during retrieval.

This

proc-

ess

could

be

viewed within

the

framework

of

stimulus sampling theory

(Estes,

1959)

if

the

stimulus elements

are

taken

to

represent

dimensions

of

organization.

For

clarity,

let

us

denote

the

image which encodes

the

cor-

rect

response

for the

current

test

as a

"c

code,"

and

denote

the

other codes

as

"i

codes."

Thus

the i

codes

are

irrelevant

codes

which should lead

to

intrusion errors,

whereas

the c

code,

if

examined, should lead

to

a

correct response. Then

the

directed

component

of

search

can be

characterized

by

the

probability that

the c

code

is in the

examination

subset, called

p

0

.

In

experiments

in

which clues

are

available

to

denote

the

organizational

dimensions

to be

searched,

/>„

may

be

close

to

1.0.

In

other situations, such

as

continuous tasks with randomly chosen

stimuli

and

responses,

p

0

will

be

lower

and

dependent

upon such factors

as the

amount

of

information

in the c

code

and its age

(where

"age" denotes

the

position

of the

code

on

the

temporal

dimension).

Although

the

breakdown

in the

directed component

can

provide

a

reasonable degree

of

forgetting,

we

shall

focus

primarily upon

the

third

mechanism

of

forgetting:

the

increasing

size

of

the

examination subset.

When searching

the

examination subset,

there

are a

number

of

possible results.

The

c

code

may be

examined

and

give rise

to a

correct response,

one of the i

codes

may be

examined

and

produce

an

intrusion response,

or

none

of the

codes

may

give rise

to a re-

sponse

and the

search terminates.

If the

search through

the

examination subset

is at

least

partially random, then

the

following

conclusions

may be

reached. When

the

size

of

the

subset

is

increased (i.e.,

the

number

of

i

codes

is

increased),

then

the

probability

of

giving

an

intrusion

will

increase,

the

aver-

age

time until

the c

code

is

examined

will

in-

crease,

and the

probability

of

giving

a

correct

response

will

decrease. When

we say

that

the

order

of

search

is

partially random,

we

mean

to

imply

that

the

order

in

which codes

in

the

examination subset

are

selected

for

consideration

may

depend upon both

the

amount

of

information

in the

code

and the

age of the

code. Clearly,

as the

amount

of

information

in a

code tends toward zero,

or

as the age of a

code increases,

the

probability

of

examining that code early

in the

search

should

decrease.

In

order

to

make

the

sequence

of

events

in

the

search clear,

an

example

of a

search

is

presented

in

Figure

2 for a

continuous

paired-associate

task.

In

this task, suppose

that

on

successive trials stimulus-response

pairs

are

presented;

on

each trial

the

stimu-

ORDER

OF

SEARCH

CODES

IN THE

EXAMINATION

SUBSET

AMOUNT

OF

STORED

INFORMATION

STIMULUS

NUMBER

TRIAL

NUMBER

1

ILi

18

33 32 31

I

1

i

I

70 69

68

67

1

30

1

66

I

1

l.l

24

29 20

I I I

65 64 63

/

•

1.

28 27

I

l

62

61

i

y-*

\

.'~^,

X

6

-•

1

-R

I.I.I

I

26 25 24 23 22 21 20 19 18 17

i

I I

* *

'

60

69 58 57

66

55 64 53 52 61

>

_

FIG.

2. A

schematic representation

of an LTS

search

in a

continuous memory

task.

PROCESSES

IN

LONG-TERM

MEMORY

189

lus

is

tested

first and

then

the

pair

is

studied.

A new

pair

may be

presented

on any

trial,

and

given several reinforcements

and

tests

at

varying

intervals.

The

bottom

row in the

figure

gives

the

trial number

for a

section

of

the

task

from

Trials

51-70.

The

number

of

the

stimulus-response pairs tested

and

studied

is

given

in the

second row.

The

third

row

shows which pairs have

had

codes

stored

in

LTS,

and the

height

of the bar

indicates

the

amount

of

information

in the

code.

The

fourth

row

indicates those codes

that

are in the

examination subset

on

Trial

70

when Stimulus

18 is

presented

for

test.

For

example,

the

stimulus

tested

may

have

begun

with

a

vowel,

and the

items

in the

examination

subset could have stimulus com-

ponents which also begin with

a

vowel.

Note

that

in

this example

the

code

for

Stimulus

18 was

stored

on

Trial

52, and

happens

to be in the

examination subset.

The

top row of the

table

gives

the

order

of

search

through

the

subset;

the first

four

codes were

examined

and

rejected

but the fifth

code

examined

was the c

code

and

enabled

the

search

to end

with

a

correct response.

It

may be

seen

how

forgetting occurs

if it is

imagined

that Item

18 had not

been tested

until

Trial

90 or

100.

In

this event, more

i

codes would

be

present

in the

examination

subset

and the

probability would

be

greater

that

an

intrusion would occur,

or

that

the

search would terminate before

the c

code

was

examined.

Interference

Various interference phenomena

are

read-

ily

predicted

by the

system. Although

in

general

the

order

of

search through

the

examination

subset will depend upon

the age

and

amount

of

information

in the

codes,

suppose

for

simplicity that

the

search order

is

entirely random.

Then

both nonspecific

proactive

and

retroactive interference

effects

are

predicted,

and in a

sense

are

predicted

to

be

equal.

That

is,

extra

i

codes

in the

examination subset added either

temporally

before

or

after

the c

code

will

cause

the

cor-

rect retrieval probability

to

drop;

in the

case

of a

random search

the

probability

de-

crease

will

be the

same whether caused

by

an

i

code preceding

or

following

the c

code.

The

drop occurs because

the

extra

codes

increase

the

amount

of

time required

to find

the c

code. Therefore,

if the

size

of the

examination

subset

is

increased,

it is

more

likely

that either response time will

run out

or

an

intrusion response will occur. Obvi-

ously

the

greater

the

degree

to

which

the

search

is

ordered temporally backwards

from

the

most recent item,

the

less

the

proactive,

and

the

greater

the

retroactive interference

effect.

Thus

if

codes

are

examined

strictly

in

temporal order,

the

average amount

of

time

until

the c

code

is

examined will

be

independent

of the

number

of

codes which

are

older

than

the c

code

and

hence

no

pro-

active

effect

will

be

expected.

One of the

best

places

to

examine nonspecific proactive

and

retroactive interference

effects

is in the

study

of

free-verbal recall

as a

function

of

list

length (Murdock,

1962).

In

this

task

a

list

of

words

is

read

to S, who

attempts

to

recall

as

many

of

them

as

possible

following

their

presentation.

The

data

are

usually

graphed

as a

serial-position curve which

gives

the

probability

of

correct recall

as a

function

of the

presentation position (and

hence

as a

function

of the

number

of

pre-

ceding

and

succeeding items

in the

list).

As

the

list length

is

varied,

the

number

of

pre-

ceding

and

succeeding items

is

systematically

varied

and it is

possible

to

apply

the

theory

to

the

resultant data.

The

application

of the

theory

is

made

particularly

easy

in

this

case

because

6"

is

trying

to

recall

all of the

list,

and

hence

the

examination subset

can be

assumed

to

consist

of all

codes that have

been

stored.

In

fact,

a

model derived

from

the

theory

has

been applied

to

free-verbal

recall data

as a

function

of

list length

and

has

proved remarkably successful

(Atkinson

&

Shiffrin,

1965,

1968a).

The

model

as-

sumes that performance decreases with list

length because more codes

are

missed

in the

memory search

at

longer

list

lengths.

This

model also predicts that some

of the

missed

codes

should

be

retrieved

if a

second recall

test

is

given following

the first.

Just

such

an

effect

was

found

by

Tulving

(1967),

and its

magnitude

is

predicted accurately

by the

model.

Three successive recall tests were

given following

a

single

list

presentation

and

only

50% of the

items recalled were recalled

190

R. M.

SHIFFRIN

AND R. C.

ATKINSON

on

all

three tests, even though

the

actual

number

recalled remained constant over

the

three

tests.

Item-specific

interference

is

also readily

predicted

via the

search

mechanism. Item

interference

refers

to

that

condition arising

when

a

stimulus originally paired with

one

response

is

later paired with

a

second, dif-

ferent

response.

In

this case

two

different

codes with

the

same stimulus

may be

placed

in

LTS.

Thus,

the

amount

of

proactive

or

retroactive interference will depend upon

the

number

of

times

the

wrong code

is

examined

and

accepted prior

to the

correct code.

In

particular,

the

degree

of

temporal ordering

in

the

search

will

affect

the

relative amounts

of

proactive

and

retroactive

effects.

That

is,

the

greater

the

degree that

the

search

is

ordered temporally backwards

from

the

most