THE TROUBLE WITH MEMORY DISTINCTIONS 
Allan Collins 
Bolt Beranek & Newman 
Cambridge, Mass. 02138 
People have an enormous variety of 
knowledge: for example, there is phonemic, 
prosodic, syntactic, lexical, semantic, and 
pragmatic knowledge; there is knowledge 
about sights, sounds, smells, tastes, 
feelings, motions, and emotions; there is 
causal, inferential, procedural, 
experiential, heuristic, and social 
knowledge, etc. Between these different 
forms of knowledge some make one set of 
distinctions, others make another set of 
distinctions, and there doesn't seem to be 
any good way to decide who is right. 
Tulving (1972), for example, has made a 
distinction between episodic and semantic 
memory, which psychologists have flocked to 
investigate. But Schank (1975) takes issue 
with Tulving's distinction, preferring to 
treat both episodic and semantic memory as 
part of a conceptual memory, which he 
distinguishes from lexical memory. Both 
Tulving's and Schank's distinctions recur 
throughout history, though with different 
labels, and are symptomatic of the problem 
of describing the global organization of 
memory. I would like to treat them as 
special cases of the more general problem of 
deciding to what extent different memory 
structures are needed for different types of 
knowledge. 
What I understand Tulving to be 
proposing is that there is an episodic 
memory, as distinct from semantic memory, 
which consists of a time-line interpretation 
of one's experiences. Semantic memory can 
be viewed as a conceptual memory, distinct 
from lexical memory so there need be no 
quarrel with Schank there. In such a scheme 
concepts in episodic memory must be linked 
to concepts in semantic memory; presumably 
to all the concepts that are processed in 
the course of understanding one's 
experience. 
Tulving's episodic memory is very much 
akin to Penfield's (1954) notion about an 
experiential record or Winograd's (1972) 
event memory, which consisted of a goal 
stack used to answer "Why" and "How" 
questions. In the case of Winograd there 
was also a semantic memory which contained 
general information about the Blocks World, 
and a visual memory of sorts in the visual 
scene he could display and manipulate. 
The Winograd example points up the 
benefits of separate memory structures. 
Because the event memory is organized 
temporally according to goals and subgoals, 
it is well structured to access information 
and make inferences about the relations 
between events. Similarly a visual scene is 
well structured to deal with spatial 
relations between objects, and a semantic 
memory is well-structured to deal with 
property relations between concepts. 
52 
The objection Schank has to the 
distinction between episodic (or event) 
memory and semantic (or conceptual) memory 
is that much of people's general world 
knowledge that would presumably be in 
semantic memory (for .example, knowledge 
about going camping) is temporal or episodic 
in structure. So why postulate a separate 
memory structure to represent it? A striking 
example is knowledge about history, which is 
general knowledge in the form of a time-line 
that does not involve one's personal 
experience. The point is that any 
representation of semantic or conceptual 
memory must have high-level concepts such as 
going camping or the Civil War. These 
concepts must include temporal structure of 
goals and subgoals of the various 
subconcepts (such as the Western Campaign in 
the Civil War) that comprise them. Thus the 
distinction that Tulving was trying to make 
can not be based on temporal structure. 
Tulving also points out a second 
important aspect of episodic memory: that it 
relates events that occur to the 
circumstances in which they occur. If one 
hears a speech such as Nixon's resignation 
speech, events that occur in understanding 
the speech are related to the physical 
circumstances at the time (aspects of the 
medium and what the listener is doing 
simultaneously) as well as a personal 
evaluation of the source (Ortony 1975). The 
knowledge in semantic memory on this view is 
abstracted from the context in which it is 
learned. For example, the personal aspects 
of the time line in an episodic memory do 
not carry over to the time line of history. 
If you read about the civil war, you may 
remember that you read about Bull Run on 
some specific page, but that information is 
part of the episodic time line and not the 
abstract historical time line. Episodic 
memory then is where our knowledge about 
experiential context is stored. 
Here again through an argument similar 
to Schank's argument with respect to 
temporal structure can be made. That is to 
say it is Just as important to be able to 
represent simultaneous context in our 
general world knowledge as it is in our 
experiential knowledge. In a representation 
of the Civil War, for example, we might want 
to relate what was happening in the Western 
Campaign to what was happening around 
Washington, and to Lincoln's feelings and 
evaluations about successive generals who 
didn't capture Richmond, etc. There is 
nothing special about representing 
simultaneous context that is peculiar to 
one's personal time-line of experience. 
There are simply no structural grounds on 
which you can distinguish semantic memory 
and episodic memory. 
In summary, people must have memory 
structure that preserves the time line of 
personal experience and its context, as well 
as structure that is abstracted from 
experience. In the brain these two kinds of 
structure may even be anatomically distinct. 
But Schank is correct in that all the power 
for representing temporal structure and 
simultaneous context must be available in 
I 
i 
I 
I 
I 
l 
I 
i 
I 
I 
I 
I 
I 
i 
I 
I 
i 
I 
both kinds of memory. This means you can't 
tell episodic memory from semantic memory by 
how the information is organized. But no 
one, not even Schank, is brave enough to 
give up the distinction completely. 
Knowledge is abstracted from experience to a 
greater or lesser extent, and it is in the 
cards that a distinction between abstract 
and experiential knowledge will keep turning 
up in one form or another. But the 
distinction is one of content, not of 
structure. 
The next question is whether Schank's 
distinction between lexical memory and 
conceptual memory is made of any sturdier 
stuff? It is based on the age old 
distinction between words and concepts, but 
raises it to a distinction between two types 
of memory. I too (Collins & Loftus, in 
press) am guilty of making the same 
distinction to account for Loftus's data 
(Grober & Loftus, 1974, Loftus & Cole, 1974) 
that response time to name objects in a 
category, behaves differently when given a 
letter descriptor than when given an 
adjective descriptor. So the distinction 
between these two types of memory at least 
has some experimental support. But the 
trouble is that the distinction melts away 
in just about the same way as the 
distinction between episodic and semantic 
memory, when you look at hard cases. 
Brown and McNiell's (1966) study of the 
tip of the tongue phenomenon showed that 
when a person can't recall a word like 
"sextant," but thinks he knows what the word 
is, he at least can tell you properties of 
the word (e.g. that it starts with "s," and 
ends with "ant/d," sounds like "secant"). 
It seems that the person is searching his 
lexical memory for the word, and that the 
information he gives about the word reflects 
the structure of the memory he uses to 
access the word. The lexical memory must be 
something like a network (or space) with 
words as nodes and ordered phonemes as 
properties. 
In contrast consider a concept like the 
barking of dogs. The concept of barking 
must of course include what we might call 
semantic or propositional properties (e.g. 
what causes a dog to bark, that barking is 
one kind of noise a dog makes). But a part 
of the concept must specify the properties 
of what barking sounds like. (Actually one 
must have a range of sounds that barking 
sounds like.) What we have then is ordered 
sounds, not unlike phonemes. So the 
conceptual memory must have the same power 
for sensory representation as the lexical 
memory. 
Similarly the lexical memory must have 
the same power for representing 
propositional information as the conceptual 
memory. For example, people must be able to 
represent that the word sextant has seven 
letters, two syllables, etc. in lexical 
memory. In these terms words turn out to be 
a special kind of concept. So if words and 
concepts must be closely interconnected and 
must have the same power of representation, 
why should they be considered to be in 
53 
separate memories? Again the distinction is 
based on content (words vs. concepts) 
rather than on structure. 
My suspicion is that if you squeeze any 
distinction between two types of knowledge 
hard enough, you will find that the 
important structural constituents are common 
to both (as was temporal structure to 
episodic and semantic memory). This does 
not negate the distinction; it only means 
that our knowledge is cut into different 
shapes from the same cloth. The same thing 
happens in the domain of living things: when 
you press the distinction between plants and 
animals with hard cases like plankton, the 
distinction breaks down. 
Part of the task of describing the 
organization of memory is to define the 
distinctions that are important and how they 
are interrelated i.e., to draw the 
organizational chart for human knowledge. 
Another part of the task is to define the 
structural elements that different memories 
share, and the structural differences, if 
there are any, that the distinctions are 
based on. The trouble with trying to draw 
the organizational chart for human memory is 
that it does not appear to have any overall 
organizing principle, such as evolution, 
that would produce a clear organizational 
structure. But then perhaps we are only 
waiting for our Darwin to come. 
ACKNOWLEDGEMENTS 
I would like to thank Mark Miller for 
his comments on a draft of this paper. This 
work was supported by a grant from the John 
Simon Guggenheim Fouundation. 
REFERENCES 
\[I\] Brown, R.W. and McNeill, D., The "tip 
of the tongue" phenomenon, Journal o__ff 
Verb~l Learning and Verbal Behavior, 
1966, 5, 325-337. 
\[2\] Collins, A.M.& Loftus, E.F., "A 
spreading activation theory of semantic 
processing." ~ Review, in 
press. 
\[3\] Grober, E.& Loftus, E.F., "Semantic 
memory: searching for attributes versus 
searching for names. Memory and 
Cognition, 1974, 2, 413-416. 
\[4\] Loftus, E.F.& Cole, W., "Retrieving 
attribute and name information from 
semantic memory. Journal of 
Psychology, 1974, 102, 1116-1122. 
\[5\] Ortony, A., "How episodic is semantic 
memory? Proceeding of the Conference on 
Issues i__nn Natural Language 
Processing, 1975. 
\[6\] Penfield, W., "The permanent record of 
the stream of consciousness,,, Proc. 
14th Int. Congr. Psychol., Montreal 
Neurological Inst., Rep. No. 486. 
\[7\] Schank, R.C., "The structure of episodes 
in memory". In D.G. Bobrow & A.M. 
Collins (Eds.) Representation and 
Understanding. New York: Academic 
Press, 1975. 
\[8\] Tulving, E., "Episodic and semantic 
memory". In E. Tulving & W. Donaldson 
(Eds.) Organization of Memory. New 
York: Academic Press, 1972. 
\[9\] Winograd, T. Understanding Natural 
Language. New York: Academic Press, 
1972. 
54 
! 
! 
! 
i 
! 
l 
! 
! 
