A widely held belief among linguists is that human language provides a window on the human mind (e.g., Chomsky 1968/2006; Pinker 1994, 2007). Under this belief or independently of it, a possible connection between grammatical structures and the organization of human memory has occasionally been suggested (e.g., Chafe 1973; Dahl 1983; Dahl to appear). For example, Dahl (to appear) claims that there is at least an indirect relationship between memory and linguistic phenomena related to time, i.e., Tense, Aspect, Mood, and Evidentiality (TAME), and “TAME categories may reflect how or from where the information expressed in a sentence is stored in the brain” (p. 3).1

The present paper is an attempt to further establish the connection between the make-up of grammar and human memory by probing a set of TAME markers that appear in realis attributive clauses in Korean, namely, -(u)n, -nu-n, and -te-n. The central claim will be that the behavior of these markers showcases how human language may encode “semantic memory” and “episodic memory” in the sense of Tulving (1972, 1983, 2002, 2005) and what kind of derivative relationship may hold between the two types of memory systems.

This paper is organized as follows: Section 2 offers a brief introduction to semantic memory and episodic memory, in particular how episodic memory differs from semantic memory. Section 3 demonstrates how the two types of memory systems can be encoded in a human language by using data that are made available from Korean attributive clauses. Section 4 concludes the paper with a brief summary and implications for future research.

1To this end, Dahl points out how sleep gives rise to “consolidation of memory”; how in numerous unrelated languages, Tense distinctions are made on the basis of what he calls “hodiernality”, i.e., referring to the day of the utterance; and how what counts as “hodiernal past” vs. “pre-hodiernal past” is determined by the time when people go to sleep in various cultures, as documented by Crane (2011) for Totela (a Bantu language spoken in Zambia).

What are known as semantic memory and episodic memory in the literature are both sub-systems of propositional memory (or declarative memory), which, along with procedural memory, constitutes human memory at the macro level.2 In Tulving’s seminal work in 1972, the two sub-types of propositional memory were believed to interact with each other at all times, but they differ in that while semantic memory is concerned with a person’s “abstract, timeless, encyclopedic knowledge” of the world that he/she shares with others, episodic memory is concerned with “unique, concrete, personal experiences” dated in the rememberer’s past.

More recent work by Tulving identifies more fine-grained differences between the two types of memory systems. For example, Tulving (1983) shows that they differ from each other in terms of (i) information, (ii) operation, and (iii) applications (or the role that memory plays in a broad range of human affairs), as summarized in Table 1. Among the differences between episodic memory and semantic memory listed in Table 1, three most notable are that (i) the reference for episodic memory is “self”, whereas the reference for semantic memory is “the universe”; (ii) that episodic memory is later developing than semantic memory; and that (iii) episodic memory is a human-specific cognitive ability, whereas semantic memory is not.

Tulving (2005) makes additional comments on the differences between episodic memory and semantic memory. He claims that only healthy humans older than 4 years old possess what he calls “autonoetic” (“selfknowing”) episodic memory, and the possession of this particular type of episodic memory allows humans to “mentally travel” both into the past and into the future through “subjective time”, whereby consciously “re-living” or “re-experiencing” things that happened in the past or “foreseeing”, “preexperiencing”, and/or “anticipating” things that may happen in the future. Other species clearly possess semantic memory, and they may even possess some type of episodic memory, but crucially, they lack the ability to mentally travel into the past or into the future with a clear mental awareness of doing so. That is, their episodic memory is not “autonoetic”, so is far less developed than what humans possess.

Yet another important property Tulving (2005) ascribes to episodic memory is that it “represents an extension of semantic memory, both in its emergence in the course of evolution and in terms of its operations” (p. 13). More specifically, in Tulving’s theoretical framework, which is known as “the serial parallel independent (SPI) model” (Tulving 1993, 1995), episodic memory is considered a derivative of semantic memory, and the two types of memory can co-exist and operate in a parallel fashion.

In more recent years, Tulving’s classification of memory has been challenged. For example, Fivush (2011) takes issue with Tulving’s (2002) definition of episodic memory, according to which, one of the two essential properties of episodic memory is memory of the specific “what, where, and when” of an experience, and the other of which is its “involvement with autonoetic consciousness”, i.e., the awareness of self having experienced the event in the past, which involves mental travel in time. Fivush claims that only the first type of memory should constitute episodic memory and the second type should be relabeled as what he calls “autobiographical memory”. His reasoning for such reclassification is that while even animals have the ability to recall specific past events including the information on the what, where, and when, only humans have the ability to recall past events involving autonoetic consciousness and notably, this ability “goes beyond the episodic memory function of guiding current and future behavior to serve social and emotional functions, including self-definition, self-inrelation, and self-regulation (Bluck & Alea 2002; Fivush 1988; Fivush et al. 2003)” (Fivush 2011: 560-561).

Despite such disagreements about exactly what counts as episodic memory and what counts as autobiographical memory, however, authors including Fivush (2011) still unequivocally agree (i) that what Tulving calls episodic memory is a human-specific phenomenon whereas semantic memory is shared by other intelligent species; (ii) that Tulvingian episodic memory involves a conscious self engaging in experiences that link the past self to current self along a personal time line; and (iii) that Tulvingian episodic memory is late developing, typically coming into existence around age 5.

Given such consensus among authors, then, we can conclude that what Tulving calls episodic memory, i.e., a memory that involves “autonoetic consciousness”, is distinct from semantic memory, and only humans possess it. And if we believe that human language provides a window on the human mind, then, we are led to hypothesize that episodic memory in the sense of Tulving is grammatically encoded, and it is likely to be encoded in ways different from semantic memory.

In what follows, I test this set of hypotheses by looking at the semantic functions of attributive clauses in Korean from the standpoint of episodic memory and semantic memory. I have chosen to examine attributive clauses rather than non-embedded clauses here because when talking about memory, humans typically talk about what they remember from the past and yet the “what” part of what they remember can be either an eventuality or some property they associate with themselves or with someone else. Hence, in order to investigate how human language grammatically encodes a Tulvigian sense of episodic memory, we need to probe an embedded clausal structure which serves as the complement of a remembering verb or as a clausal modifier of an individual-denoting expression, such as the attributive clause construction in Korean.

2Propositional memory systems differ from procedural memory systems in several respects: e.g., (i) information handled by the former has a truth value, whereas information handled by the latter does not; (ii) information retrieved from the former can be contemplated introspectively or attended to internally, whereas that retrieved from the latter cannot; and (iii) propositional knowledge about something can be communicated to others through language or other methods, whereas procedural knowledge can only be demonstrated through highly specific behavior (Tulving 1984: 224).

I begin this section by briefly introducing some of the basic properties of Korean and its attributive clause construction that we will be looking at.

Korean is morphologically an agglutinative language and syntactically a Subject Object Verb (SOV) constituent order language (Sohn 1999). Scrambling is possible but the language exhibits strict head-finality (e.g., the predicate always occurs at the end of a sentence).

Due to its agglutinative morphological characteristics, attributive (ATT) clauses in Korean are comprised of a verb stem and a relative clause (RC) marker, both of which are bound morphemes (although one is a right-bound morpheme and the other is a left-bound morpheme).

ATT clauses in Korean may or may not contain a gap. If they contain a gap, then they instantiate externally-headed relative clauses (RCs), which contain an empty category that is co-indexed with the head noun (i.e., the noun that an RC modifies). If they do not contain a gap, then, they instantiate gap-less noun-modifying clauses or nominalized clauses that occur in Subject or Object position.

ATT clauses in Korean can bear either realis or irrealis mood (Lee 1993; Sohn 1999), and those bearing realis mood will be our main concern here.Realis ATT clauses in Korean end with the adnominalizer/RC marker -(u)n3 and, between the verb stem and the RC marker, additional TAME markers may occur, namely, -nu-, -te-, and -ess-te-. To see this, consider (1), where the first four data illustrate RCs and the last illustrates a gap-less ATT clause. (Here and below, ‘e’ stands for a gap or an empty category that is co-indexed with the head noun of an ATT clause.4)

As indicated by the English translations of the above data, each of the grammatical markers that occur between the verb stem and the RC marker in Korean contributes some sort of aspectual semantics, i.e., they specify the relation between the situation time and the topic time. More specifically, in (1a), the null marker Ø indicates that the situation at hand temporally precedes the topic time; that is, it behaves as if it carries the semantics of an anterior or perfective marker. In (1b), the morpheme -nu- indicates that the event of someone leaving is in progress at the topic time, so it seems to carry the semantics of a non-past progressive marker. In (1c), the same event was on-going at some time before the speech time, so the morpheme-te- seems to carry the semantics of a past progressive marker. Finally, in (1d), the event was completed before some past time, so the complex marker -ess-te- seems to contribute a pluperfect (i.e., ‘past in the past’) meaning here.

A closer examination quickly reveals, however, that the four temporal markers just introduced are not just concerned with Aspect or Tense; rather, they have something to do with Mood, a grammatical category that is concerned with propositional attitudes or knowledge ascriptions (Aikhenvald 2004), and Evidentiality, a grammatical category that has to do with indicating informational source (Chafe & Nichols 1986; Aikhenvald & Dixon 2001). And this is particularly true of -nu- and -te- that co-occur with the RC marker. Since it is well established that -nu- and -te- behave differently when they occur as part of the RC marker as opposed to as part of the verb cluster of a non-embedded clause (e.g., Lee 1993, Lee 2012), here and below, I treat -nu-n and -te-n as complex clusters, rather than separating -nuand -te- out from -(u)n.

To illustrate the modal or evidential nature of -nu-n and -te-n, consider first (2), which contains -te-n. In (2), -te-n has the semantics of a past habitual marker, rather than a past progressive marker, unlike in (1c). And interestingly enough, depending on context, this sentence may or may not be uttered felicitously: it can be uttered felicitously in the contexts given in (3) but not in the context given in (4). To verify this judgment, I conducted a survey with 12 native speakers of Korean (6 males and 6 females), and eight of them said that they can only accept sentence (2) in context (3a), two said that they can accept it either in (3a) or (3b) but not in (3c), and the remaining two said that they can accept it in all three contexts including (3c) (and there was no correlation between gender and the informant’s permissiveness on (2)).

Although there is inter-speaker variation on the judgment, what matters for our present purposes is that no speaker judges sentence (2) to be acceptable in the context given in (4) and this shows that -te-n can be licensed only if the speaker has reliable evidence for the situation described by the ATT clause containing it.6 Further evidence for the evidential nature of -te-n comes from the fact that if the speaker does not have a truly reliable source to attribute the property described by the -te-n ending clause to the denotation of its head noun, then, he/she has to use a completely different ATT clause ending which involves a hear-say marker, as shown in (5).

Turning now to illustrating the modal or evidential nature of -nu-n, typically, ATT clauses ending with -nu-n describe situations that are on-going at the topic time, as shown in (1b), but they can also attribute a factual property to the head noun’s denotation based on what the speaker knows about the world. To see this, consider (6). Here, the content of the RC marked by -nu-n has not happened as yet, so it cannot be the case that the speaker directly observed it or someone or something she completely trusts has informed her about its occurrence. But from what the speaker knows, the event under description is planned to occur tomorrow, so it counts as a fact of the world.

Notice that (6) cannot be followed by the speaker’s hedging about the veridicality of the propositional content of the RC, as shown in (7), and in order to deny its truthfulness, a hearsay ending has to be employed, as shown in (8). Importantly, however, even in hearsay contexts like (8), the ATT clause is marked by -nu-n (as is the case in (5)), and this is because according to the speaker’s knowledge of the world, what is quoted in the form of an embedded clause is a fact.

Even this cursory introduction to the realis ATT clause markers in Korean shows that they carry more than one type of grammatical meaning, and their semantics has something to do with Mood (i.e., knowledge state) and Evidentiality (i.e., information source), as well as Tense (i.e., relation between speech time and topic time) and Aspect (i.e., relation between situation time and topic time). Therefore, we can call them TAME categories. In the next three subsections, I show how they divide the work in linguistically encoding episodic memory and semantic memory.

When we reexamine ATT clauses in Korean in the light of the definitions that Endel Tulving offers for semantic memory and episodic memory, there are several reasons to think that some ATT clauses comprised of a verb stem and -te-n such as (1c) encode episodic memory for the past.

First of all, according to Tulving (1983), episodic memory for the past is something that people report as what they “remember” whereas semantic memory is something that they report as what they “know” (see Retrieval Report under Operations in Table 1), and notably, ATT clauses ending in -te-n can occur as the nominalized complement to the verb kiekha-ta ‘to remember’ but not as the nominalized complement of the verb al-ta ‘to know’. To see this, compare (9a) and (9b).

Secondly, when uttered out of the blue, -te-n-ending ATT clauses require spatio-temporal adverbials, as exemplified in (10), and this shows that they describe episodic events or particulars, viz., concrete instantiations of an event type. Importantly, this property is reminiscent of how experimental queries testing episodic memory are in the form of ‘when’ (i.e., time?) or ‘where’ (i.e., place?) so that they can set the scene for the situation that is being recalled (see Retrieval Queries under Operations in Table 1).

A similar phenomenon is observed by the contrast between some ATT clauses that end with -(u)n and some that end with -te-n. To see this, compare (11) and (12). In (11), -te- is not allowed because the speaker is asking the hearer about how many people he/she has dated thus far, so what is at issue is not some particular eventuality that is anchored at some particular spatio-temporal location. By contrast, in (12), -te- is required and this is because, here, what is at issue is some particular eventuality.

Thirdly, when a -te-n-ending clause reports on what the speaker personally observed in the past, its informational source is “sensation” and its reference is “self” (i.e., the person who is doing the remembering of past eventualities), which are two of the defining properties of episodic memory that Tulving (1972, 1983, 2005) identifies. To illustrate, reconsider (2). When uttering (2) in contexts like (3a), as a Korean native speaker, I intuit that I mentally travel back to the times when I observed John’s living in some contextually salient house that we are looking at. Similarly, in order to answer the question in (12) with the -te-n marking on the ATT clause, I must go back to the time that the speaker and I are talking about and count the number of people I was dating at that time. Interestingly, in order to respond to (11), I must also go back to some past times, but unlike the case with (12), I do not have to remember the events as if I were re-experiencing them; that is, I do not undergo sensation. Instead, for cases like (11), I only need to remember the events that instantiate my dating someone and count how many such events have existed in my life and how many individuals they have involved.

Fourth, -te-n can select for both adjectival and verbal predicates, but not every predicate may co-occur with it because what a -te-n ending clause describes must be recallable. To see this, compare first (13) and (14). In the case of (13), the head NP’s referent is in the 1^st person, so we can construe each datum in such a way that the speaker is recalling her own past experiences, namely, some emotive/cognitive states that she was in at some time in the past. In the case of (14), on the other hand, the head NP’s referent is in the 2^nd or 3^rd person, and since the speaker has limited access to another individual’s internal states, the entire NPs do not come out grammatical in the given discourse context.

Notice that the data in (14) improve in grammaticality if the embedded predicates co-occur with verbs of ‘seeing’ or if they are turned into verbs, as illustrated by (15). Such changes can improve on the grammaticality because they make the embedded clauses describe properties that are observable to the speaker, which can be stored in episodic memory and can later be retrieved.

Additional evidence for the strong involvement of self in uttering RCs ending with -te-n comes from the fact that typically, they cannot occur in sentences that have non-first person sentential subjects. To see this, compare the sentences in (16). The two sentences are identical except for the subject’s person feature (i.e., first person vs. third person), but while (16a) is grammatical, (16b) is plain ungrammatical in the given discourse context. Notably, when uttered by the narrator of a novel or a documentary film, (16b) can be judged fine but in such cases, the narrator is assumed to have access to the subject’s mental state, so can represent the sensation that the subject experiences in recalling the eventuality described by the embedded clause.

At this juncture, I should note that although remembering verbs in Korean such as kiekna-ta ‘to recall/remember’ and sayngkakna-ta ‘to recall’ may select for nominalized clauses that do not contain -te-, such clauses do not engender an interpretation in which the speaker vividly recalls some past eventualities as if she were re-experiencing or re-living them. To see this, consider (17). In these sentences, what is being conveyed is that the speaker simply remembers some past eventuality or the fact that p (where p stands for a proposition). That is, uttering such sentences is devoid of the involvement of self in the event described, i.e., what Tulving calls “autonoetic consciousness”.

Let me now point out that the propositional content of some -te-n ending clauses is verified by looking at what the speaker remembers about the past, independently of what is true in the actual world, and this is in line with what Tulving (1983) claims about episodic memory, namely that the veridicality of episodic memory is determined by one’s personal belief about his/ her past experiences, regardless of what the actual experiences were like or what other people report about them. By way of illustration, suppose that I say (18) but what I remember is inaccurate because I suffer from dementia (for example, in the actual world, Mina used to run a shoes store, not a clothing shop). Even in such contexts, however, one cannot refute the content of what I remember, although they may say that (due to illness), I have an inaccurate memory of (some) past eventualities.

While Korean employs -te-n as a way to encode episodic memory for the past, as a way to encode episodic memory for the future, it employs either -(u)n or -nu-n, and the choice between the two markers is based on the semantics of the lexical predicate at hand. In brief, if the predicate has adjectival semantics, it combines with -(u)n; if it has verbal semantics, then it combines with -nu-n (Sohn 1999; see Lee 1993 for a slightly different characterization).

To see this, compare first the data in (19). In both cases, the embedded clause describes an eventuality that exists in some possible world and talking about it involves the speaker’s “mental travel into the future”, a property that characterizes episodic memory (Tulving 2005). Given this, we can conclude that both -(u)n and -nu-n can encode episodic memory for the future.

Consider now the paradigm in (20): (20a) shows that the adjectival predicate ttokttokha-ta ‘to be smart’ is not compatible with -nu-. On the other hand, (20b) shows that the verbalized predicate ttottokhayci-ta ‘to become smart’ requires the morpheme -nu-. Taken together, these facts show that the bare realis RC marker -(u)n and its more augmented variant -nu-n have different selectional properties, even though both can provide ways to grammatically encode episodic memory for the future.

Interestingly enough, in order to encode semantic memory, Korean utilizes the same set of TAME markers that it uses for encoding episodic memory for the future, namely, -(u)n and -nu-n.

First of all, there is ample evidence that -nu-n can encode semantic memory. The first piece of evidence comes from the fact that ATT clauses ending with -nu-n may describe generic, habitual properties of individuals that obtain at the speech time. To see this, consider (21). Notice also that when a nu-n-ending clause describes a generic property, its generic nature cannot be denied, as evidenced by the oddity of the discourse in (22).

The claim that -nu-n marking provides a way to describe generic properties or timeless facts in the world receives further support from the fact that unlike ATT clauses ending with -te-n, those ending with -nu-n do not require spatio-temporal adverbials, as illustrated in (23) (modulo the semantic difference between cases where such adverbials are present and cases where they are not, as can be seen by comparing (23) and (24)).

Furthermore, ATT clauses ending in -nu-n can occur as part of ‘what’ questions which inquire about general properties of things in the world, as shown in (25), or in answer to such questions, as shown in (26)-(27). Importantly, this is reminiscent of what Tulving (1983) observes about semantic memory, namely that experimental questions testing semantic memory are in the form of ‘what’ (see Retrieval Queries under Operations in Table 1).

Another notable characteristic of -nu-n ending clauses such as those in (25)-(27) is that what they describe must be supported by social consensus. That is, their veridicality is something that is determined by an entire speech community rather than by a single individual, unlike the case with -te-n ending clauses. Importantly, such property is analogous to the way in which semantic memory behaves according to what Tulving (1983) claims.

Finally, ATT clauses ending with -nu-n exhibit the opposite behavior to ATT clauses ending with -te-n in terms of their ability to occur as the complements of certain propositional attitude predicates. In short, they can occur as the nominalized complement to the verb al-ta ‘to know’ but not as that of kiekha-ta ‘to remember’, as shown in (28). And this accords with Tulving’s (1983) claim that semantic memory is something that people report as what they “know”, rather than what they remember (see Retrieval Report under Operations in Table 1).

Turning now to investigating when -(u)n may be employed to encode semantic memory, it turns out that it is employed whenever -nu-n cannot be, namely, (i) when the lexical predicate at hand is adjectival or (ii) when the semantic memory at hand is derived from some past eventuality. To see this, consider first (29). These data show that adjectival predicates are incompatible with -nu-n, so in order to describe timeless, semi-permanent, or generic properties of individuals, they must co-occur with -(u)n.7

Given the generic semantics of the data in (29), it is not surprising that ATT clauses marked by -(u)n can occur as part of ‘what’ questions which inquire about the general properties of things in the world or in answer to such questions, as shown in (30)-(32), in a manner analogous to (25)-(27).

To illustrate the contexts in which an ATT clause ending with -(u)n encodes semantic memory that is derived from past eventualities, consider (33)-(34). The lexical predicates constituting the ATT clauses here are all verbal predicates and yet they co-occur with -(u)n, rather than, -nu-n. This is because here, each ATT clause describes a timeless, encyclopedic knowledge of the world that (almost) any Korean native speaker older than age 10 possesses, based on what they have learned from school or whatever relevant sources that deal with Korean history.

Before closing this section, let me briefly remark on the semantics of -te-n and grammatical encoding of semantic memory in Korean. Based on the fact that sentence (2), repeated below, can be uttered in a context where John is a famous historical figure (i.e., context (3b)), one may suspect that just like -(u)n, -te-n may also be employed to encode semantic memory which is derived from a well-known past eventuality or a proven historical fact.

While such reasoning is clearly logically based, I do not think that -te-n has the same status as -(u)n even when it occurs in a sentence like (2) and is uttered in a context like (3b). One reason for drawing this conclusion is that as noted above, most Korean speakers are hesitant about accepting (2) in such contexts: recall that only 4 out of 12 speakers I have consulted judge the sentence to be acceptable in (3b). Furthermore, even those who can accept (2) in (3b) have reported to me that they prefer variants such as those given in (5) and (35), which contain hear-say markers and thus allow the speaker to remain non-committal as to the truthfulness of the propositional content of the embedded clause.

Finally, unlike -(u)n ending clauses, -te-n ending clauses do not make a good candidate for the template we have been using to test semantic memory, i.e., the question-answer pairs such as those given in (33)-(34). To see this, consider (36). This question-answer pair has the same structure as (33), but here, -te-n is used instead of -(u)n and the answer is judged to be ungrammatical, despite the fact that (almost) any Korean speaker older than 10 years knows that General Swunshin Lee protected Korea when the Japanese invaded the country during the Chosun Dynasty. The ungrammaticality of the answer in (36) is unlikely to be due to the past imperfective semantics of -te-n because during Imcinwayran, there were numerous battles between Koreans and the Japanese and during those times, General Lee continuously fought for Korea.

What might be the reason for the unsuitability of sentences like (36A) as answers for questions like (36Q), which inquire about semantic memory? I conjecture that their unsuitability is due to the evidential semantics of -te-n. More concretely, because of its evidential nature illustrated in section 3.1., -te-n can be felicitously employed only when the speakers wishes to inform the hearer what she knows about the world based on the collection of evidence she has at her disposal (or what she has personally observed or experienced). This line of reasoning receives support from the fact that (2) can be uttered by a historian who has direct evidence to assert that the house under discussion is where John used to live but even in such contexts, the knowledge that she is sharing with the hearer is not solidly established as an encyclopedic knowledge in the speech community as yet, so it still counts as a more or less personal knowledge.

3The realization of the vowel “u” in the RC marker -(u)n is determined by whether the preceding stem ends in a consonant or not. If it ends in a consonant, then “u” is realized; if not, the RC marker surfaces as -n.  4In transcribing the Korean data presented here, I adopt Yale Romanization (Martin 1992), following standard practice in Korean linguistics. And for the interlinear glosses, the following abbreviations are used. But I do not gloss the morphemes -nu- and -te- here because there is a lack of consensus among authors concerning their exact syntactic/semantic status. Concerning the syntax/semantics of -nu-, see HS Lee 1991, 1993, Ko 2007, and references therein; concerning -te-, see Yang 1972, HB Choi 1983, Huh 1987, HS Lee 1993, KD Lee 1993, Ko 2007, KS Chung 2007, 2010, and C Lee 2012. Acc: accusative case; Ant: anterior aspect; CL: classifier; Conn: connective; Cop: copula; Hon: honorific marker; Imprf: imperfective aspect; Infml: informal style; Ind: indicative sentence ending; Loc: locative; Nom: nominative case; Nml: nominalizer; N.Pst: non-past tense; Pass: passive voice; Prf; perfective aspect; Pst: past tense; Q: question ending; Ouot: quotative marker; Rel: relativizer; Rr: realis mood; Top: topic.  5In Korean, ATT clauses and non-ATT clauses have different TAME systems although they seem to employ similar morphemes. To illustrate, while ATT-clauses employ the null morpheme Ø to express temporal precedence in non-stative contexts, in the same contexts, non-ATT clauses employ the suffix -ess-, as shown in (i). In addition, when occurring in ATT clauses, the morpheme -nu- is compatible with -ko iss- but when occurring in non-ATT clauses, it is not. To see this, consider (ii).   6In the literature, the evidential nature of the single morpheme -te- has been much discussed but the evidential nature of the complex cluster -te-n has not been, and authors tend to believe that only -te- (i.e., without the RC marker attached to it) carries evidential semantics. For discussion, see KS Chung 2012 and C Lee 2011/2012.  7The reason why adjectival predicates cannot co-occur with -nu-n is most likely to be due to the imperfective aspectual semantics of -nu- (Sohn 1999). For reasons not fully understood as yet, under all circumstances, Korean adjectives cannot bear imperfective aspect marking, unlike English adjectives, which are compatible with imperfective morphology although what bears the imperfective morphology in English is the copular verb (e.g., John is being kind; You’re being silly.). To see this, compare (i) and (ii), which respectively exemplify imperfective aspect marking on verbal predicates and that on adjectival predicates in Korean.

One of the main findings of this paper has been that Korean employs the ATT clause marker -te-n as a way to encode episodic memory for the past and -(u)n and -nu-n as a way to encode episodic memory for the future and whenever ATT clauses ending with such markers encode episodic memory, uttering them involves “autonoetic consciousness” in the sense of Tulving (2005). Another important finding has been that -(u)n and -nu-n can “double-up” as semantic memory encoders and the choice between them is made by the semantic properties of the lexical predicate at hand and the source of the semantic memory at hand. When encoding a semantic memory that is derived from observing recurring eventualities (e.g., properties of some particular species or habits of individuals), -nu-n is employed for verbal predicates and -(u)n is employed for adjectival predicates; when encoding a semantic memory that is based on historical facts, -(u)n is employed regardless of the predicate type of the ATT clause. These findings are summarized in tables 2 and 3 .

What has been uncovered here shows that what Tulving calls episodic memory can be grammatically encoded, and this partly confirms our initial hypothesis outlined in section 2. That is, given the human-specificity of “autonoetic consciousness”, a memory system that involves it will be linguistically encoded. In addition, our finding that Korean utilizes the same set of TAME markers to encode both semantic memory and future-oriented episodic memory lends support to Tulving’s theory of memory even though it is slightly different from what we initially anticipated. Recall that according to Tulving (1993, 1995), episodic memory is an extension of semantic memory and hence the two types of memory can co-exist and operate in a parallel fashion. Given such relationship between semantic memory and episodic memory, if we posit that grammatical structures reflect the organization of human memory, then we can conceive of grammatical morphemes which primarily encode semantic memory but which may also encode some of its extensions, namely, episodic memory. And this line of reasoning accounts for the initially unexpected versatility of -(u)n and -nu-n observed here.

While these two findings are largely in agreement with our own predictions or predictions made by the mainstream theory of memory, the way in which Korean grammatically encodes episodic memory raises several questions that may require rethinking some of the prevailing views on memory, language, and thought.

First of all, the fact that Korean employs -te-n to encode past-oriented episodic memory but employs -nu-n and -(u)n to encode future-oriented episodic memory raises the possibility that the two types of episodic memory may be qualitatively different, unlike what Tulving’s (2002, 2005) analysis seems to assume.

Secondly, the fact that Korean morphologically conflates semantic memory encoders with future episodic memory encoders can be taken to suggest that episodic memory for the future may be more early developing than episodic memory for the past.

Thirdly, given the well-established finding in the psychology literature that even animals have semantic memory, the fact that Korean utilizes the same set of morphemes to encode semantic memory and future-oriented episodic memory lets us hypothesize that even animals may have a rudimentary level episodic memory for the future and what is truly uniquely human may just be the ability to recall past autobiographical eventualities with a strong sense of self-involvement and sensation, to take a somewhat extreme stance on the matters here.

Next, the fact that Korean has a grammatical way to encode past-oriented episodic memory and this morpheme is formally distinct from other memory encoders leads us to conjecture that Korean speakers may have a more remarkable ability to retrieve episodic memories from their youngest years, compared to speakers of languages that lack such grammatical devices.

The presence of an episodic memory marker in Korean also lets us predict that a relatively small number of Korean speakers will suffer from childhood amnesia, in comparison to speakers of languages that lack a grammaticalized episodic memory marker.8

If these predictions are borne out, then they will also suggest that there is an intimate (causal) relation between language and thought—that is, language does shape thought, as held by advocates of linguistic relativism.

Finally, our finding that Korean has a grammatical way to encode autonoetic episodic memory makes an interesting prediction on language acquisition: Based on a series of experiments investigating children’s acquisition of Evidentiality in Korean (a language with overt evidential morphology) and English (a language without overt evidential morphology), Papafragou et al. (2007) have claimed that contrary to relativistic expectations, children’s ability to reason about information sources is developed independently of the properties of the exposure language, i.e., regardless of whether it has overt evidential morphology or not. But the findings reported here make us wonder whether similar results will also obtain for an experiment that inquires about Korean children’s comprehension and production of -te-n and their ability to retrieve episodic memories or specific autobiographical events.

It is hoped that future research provides illuminating answers to the questions addressed here. Another question that I have to leave for the future is whether other languages possess evidential markers that are comparable to -te-n and if so, whether such markers perform a similar function of encoding episodic memory in the sense of Tulving (2002, 2005).

8I thank Roman Taraban for helping me to see this possibility.