Memory: Irreducible, Basic, and Primary Source of Knowledge.

Memory: Irreducible, Basic, and Primary Source of Knowledge 1.1 Anti-Reductionist Argument This paper shows how memory generates knowledge. It argues against preservationism, the epistemic doctrine that memory can at most preserve knowledge generated by other basic sources. (Lackey 2011, 317) Since preservationism claims that memory cannot generate knowledge, it is committed to the reduction of knowledge derived from memory to one or more of the other basic sources of knowledge: perception, reason, or introspection Preservationism may accept testimony as a fourth basic source of knowledge. However, I am not familiar with a philosopher who held a preservationist view of memory and considered testimony an irreducible basic source of knowledge.. (Plantinga 1993; Dummett 1994, 262; Audi 1997, 410) I show here how memory can and does generate knowledge that is irreducible to other basic sources of knowledge. My argument that memory, a type of basic source of knowledge, can generate new knowledge and not just preserve it rests on showing how multiple token memories together generate new knowledge --memorial knowledge. Other anti-reductionists, or “generativists,” have argued for equality and cooperation between the five basic sources of knowledge. As P. F. Strawson (1994) put it, perception, testimony and memory form “a community of equals” irreducible to each other. For example, memory and perception participate in the generation of scientific knowledge since scientists must recall their experiments before writing them down. “The common omission of this topic from epistemological studies is a bit of a scandal.” (Lewis 1946, 333) My argument is more radical than Lewis’ and Strawson’s epistemic “egalitarianism” in arguing that in some epistemic contexts memory generates knowledge by itself or with just trivial assistance from other basic sources of knowledge. Memory can be a primary basic source, first among equals. After clarifying the conceptual distinction between primary and trivial basic sources of knowledge, I outline an ontology of information transmission from events to memory as an alternative to causal theories of memory. I derive from information theory a concept of reliability of memories as the ratio of retrieved information to transmitted information. I distinguish the generation of knowledge from reliable memories from the generation of knowledge from unreliable memories. Reliable memories can generate knowledge through narrative and colligation. Unreliable memories can generate knowledge if they are coherent and independent and the prior probability of the knowledge they generate is sufficiently low or high. Possible confounders and the difficulties in ascertaining the independence of memories challenge the generation of knowledge from unreliable memories. It is possible to overcome these challenges, especially when memories in different minds are primary basic sources of memorial knowledge. The obvious conclusion of this article is that preservationism and epistemic reductionism of memory are disproved. 1.2 Primary and Trivial Basic Sources of Knowledge Traditionally, epistemology has distinguished, at most, five basic sources of knowledge: empirical, from perception; a-priori, from reason; testimonial, from what other people say; self-knowledge, from introspection; and memorial, from remembering. The orthodox concept of “basic” source of knowledge proposes that it “yields knowledge without positive dependence on the operation of some other source of knowledge (or of justification).” (Audi 2002, 72) Anti-reductionist epistemologists have argued that most knowledge results from corroboration and cooperation between several of these basic five types of sources. To borrow P. F. Strawson’s (1994) example, when we gain knowledge from reading, we use our senses to perceive the page, we learn from the testimony of the author, and we rely on our memory of the language of the text to semantically decode the words and understand the text. Strawson (1994) and Lackey (2008) argued that much of our knowledge has multiple types of sources that are sometimes basic, irreducible to each other. I argue that when multiple types of sources generate knowledge, the sources are often not equally significant. For example, in Strawson’s example, though empirical and memorial basic sources are involved in the generation of knowledge from reading, the author’s voice, a testimonial source, is more important than the other basic sources. When we form a belief on the basis of reading we assume trivially that our senses and memory are reliable. Epistemically, the testimony of the book is more questionable. If a belief we formed on the basis of reading is defeated, we would question the author’s testimony rather than our senses or memory. I call the primus inter pares, first among equals, basic source of knowledge the primary basic source. I designate less significant basic sources that participate in the generation of knowledge trivial sources. Primary basic sources may be less reliable or of indeterminate reliability, or may be less embedded in our web of beliefs than trivial basic sources. Surrendering trivial basic sources of knowledge would be epistemically “costlier” than giving up on primary basic sources, measured in consequent epistemic work to repair webs of beliefs. For example, a web of beliefs will not be as affected by concluding that the testimony of an author is unreliable as when admitting that one’s senses or memory are unreliable. A simple method for distinguishing primary from trivial basic sources of knowledge is to assume a defeater to the belief generated by several basic sources and intuit which of the sources would come first under epistemic suspicion and be the most likely to be eliminated first consequently. For example, suppose somebody asks Winston how many armchairs he has at home. Winston does not retrieve a single memory of all his armchairs. His basic epistemic assumptions are: He recalls two armchairs in his living room. (memorial source) He recalls one armchair in his study. (memorial source) He does not recall armchairs anywhere else at home. (memorial source) He reasons that 2 + 1 + 0 = 3 (a-priori source) Therefore, Winston knows that he has three armchairs at home. (1)-(3) are primary basic memorial sources, whereas (4) is a trivial basic a-priori source. Suppose the belief is defeated, e.g. the armchairs are counted and there are more or less than three. Obviously, at least one of the above four assumptions must then be false. We would suspect first that one of the primary basic sources (1-3) is false and not the trivial basic source (4). Eliminating (1)-(3) would be epistemically less costly than doubting basic arithmetic. Winston’s knowledge of how many armchairs he has at home is based primarily on memory; it is memorial. The distinction between primary and trivial basic sources of knowledge allows me to argue that some knowledge is based primarily on memory and trivially on other basic sources. 1.3 Memory and the Flow of Information All events transmit information signals. Memories receive few of them. Memorial knowledge results from retrieving some information signals from memory. Between the transmission of the information and its eventual retrieval from memories, information is encoded (the form of mental encoding does not concern us here) but not expressed in a period of latency. Retrieving the information requires decoding (its form also does not concern us here). During the period of latency signals can be mixed up with “noise.” Between the transmission of information from events and its retrieval, the information may or may not be experienced and may or may not be believed or constitute knowledge. Information may flow directly to memory without being consciously experienced or believed. It may become knowledge only once it is retrieved, or as I show later, after several memories are retrieved. Shannon’s classical information theory interprets information as reduction in uncertainty. Shannon codified his understanding of information in the Inverse Relation Probability Principle: the more probable or possible ρ is, the less information it is carrying. The more improbable the signal, the more information it is transmitting. For example, ceteris paribus memory of the outcome of a lottery draw conveys more information than memory of the result of a coin toss because it has a lower prior probability and eliminates more alternatives than the first. (Cf. Dretske 1981) Some information flows and memories are respectively causes and effects; and some causes of memories transmit information. But though the causes of memories and the origins of the information signals they receive partly overlap, they are not identical. Causes do not have to transmit information to effects. For example, in the first volume of Proust’s Remembrance of Things Past the narrator’s madeleine cookies and herbal tea cause him to remember lost times but they do not transmit information to those memories. Causes do not have to encode information, and effects do not have to decode it. Therefore, some causes and effects, unlike events and their memories, can be simultaneous. Conversely, it is possible to know that some memories receive information from the past without knowing the causes of the information flow, for example, knowledge based on semantic memories, such as memories of our native tongues, or “bare factual memory,” facts we remember without knowing how or why we acquired them. (Williamson 2007, 110-111) We have no episodic memories or other knowledge of the causes of our acquisitions of our native tongues or bare factual memories. Further, an epistemology of memory does not have to assume that past events cause present memories. For example, Plato’s theory of knowledge distinguishes memory as a receiver of transmission of information from the past from causation in suggesting that some information flows to memory have no known causes. Socrates and Plato held that we are reminded of, rather than learn, much of what we know. Plato offered in the Phaedo a myth to explain how souls remember without being taught. But Plato interjected myths into philosophical exposition when he had no philosophical argument for an opinion and may have wanted to draw the readers’ attention to it. A Platonic epistemology of innate ideas (think of an idealized and expanded version of Fodor’s nativism that considers all lexical concepts innate) may accept Plato’s position on knowledge and memory without taking seriously the myth about the transmigration of souls. Such a position does not have to offer causes of remembered knowledge if it can prove that we remember without being taught. An epistemology of memory that is founded on an ontology of information flows rather than causal chains avoids the kind of counterexamples that have been developed against causal theories of memory and that imagine grotesquely unreliable causal processes that generate through epistemic luck, coincidentally but consistently, true memories. Any viable causal theory of memory must add “suitable” conditions to avoid being too broad. (Bernecker 2010, 128-154; Michaelian 2016, 75-96) Some suggest the addition of conditions about information, how it is transmitted to memory and whether memorial knowledge at the end of the process “matches” or “fits” in some sense the source of the information. Concentrating exclusively on information and its transmission from event to memory and thence to knowledge without elaborating on possibly redundant causation, representation, experience, traces-engrams, content, fits and so on is more parsimonious than causal theories of memory yet dodges the obvious counterexamples. In information theory, reliability is the ratio of received information at the end of the process of information transmission to the information transmitted at its origin. For example, high fidelity sound has high reliability because its listener receives most of the auditory information transmitted at the time the sound was generated with little noise. Reliability can then be a predicate of both the process of information transmission from past events to present memories and of memories. The sources of the information signal that memories receive do not have to be past experiences or beliefs; information flows can pass below the threshold of consciousness without forming experience or belief. By comparison, process reliabalism measures the reliability of an epistemic process by the ratio of true beliefs to total beliefs generated by that process, ignoring information lost in the process. Michaelian (2016, 40) argues that it is impossible to evaluate the reliabilities of processes that generate beliefs from memory because they would have to rely on memory itself for evaluating which beliefs are true. Michaelian overstates his case since memory does not have to epistemically bootstrap itself when it is possible to measure the coherence, or what Lewis (1946) called “congruence,” of beliefs generated by types of memory with knowledge generated by other basic sources of knowledge. But since it is impossible to measure the frequency of true beliefs among all the beliefs that types of memory have generated, it would be necessary to measure the frequency of true beliefs in samples of them. It may be difficult in avoid unintentional sampling biases. The concept of reliability that information theory developed is more useful for the epistemology of memory because it can rely on information theories or generalization about the transmission of information to memory in various contexts, e.g. that ceteris paribus memories from early childhood are less reliable than later memories or that ceteris paribus the older memories are, the less reliable. The parts of common law that deal with admissibility of testimony and the competence of witnesses whose testimonies rely primarily on memory assume such folk psychological information theories of memory. Psychology can offer more rigorous and precise theories about the transmission of information from events to retrieved memories. This paper is divided into two parts, respectively showing how reliable and unreliable memories, in this information theory sense, generate knowledge as primary basic sources of knowledge. 2.1 The Generation of Knowledge from Reliable Memories Before proceeding to show how reliable memories generate knowledge, I should mention two other types of generation of knowledge from memory that are discussed in the literature. Even preservationists acknowledge that memory can receive information flows that do not form doxastic states like beliefs, and later generate knowledge from them. Bernecker (2010, 99) dismisses such cases under the rubric of “inattentive remembering.” He claims that despite the absence of belief at t1 when the memory is formed, there is a “representation” at t1 that has identical epistemic status to that of the belief that would be formed from memory. “What distinguishes S's propositional attitude at t1 from his propositional attitude at t2 is awareness or conceptualization but not justification.” However, there is no reason to believe that when information flows through the senses to memory, bypassing consciousness, it creates a representation, let alone a justified one. Information is encoded in memory and later decoded. When it is encoded, it does not have to constitute a justified representation and when it is decoded it can serve as justification for new knowledge or belief. In between, memory in its encoded latent state is not a representation. (Senor 2007, 199) More controversial is Lackey’s (2008, 251-277) argument that memory can generate knowledge when a previously undefeated defeater that prevented the generation of knowledge from memory is defeated, for example, by new evidence. A defeater can be normative, if it implies that the belief ought to be defeated, or a psychological state of mind that prevents the agent from knowing, like a trauma. (Lackey 2008, 44-5) More orthodox distinctions are between “external” propositional or factual defeaters, and “internal” states of mind. (Cf. Grundmann 2011) The debate that ensued (Senor 2007; Lackey 2007) revolved around the meaning of preservationism, the relationship or lack thereof between memory and justification, and the analysis of conflicting and inconclusive intuitive interpretations of thought experiments. I excuse myself from joining this debate because if Lackey’s argument is granted, moderate preservationism may claim that memory usually or mostly just preserves knowledge except for rare cases of generation of knowledge when a defeater is defeated, while denying radical generativism that argues that memory can and often does generate new knowledge. (Bernecker 2010, 98-99) I will defend radical generativism by showing how common, ordinary, and simple cases of generation of new knowledge from memory can satisfy even Senor’s onerous conditions for memory to be epistemically generative: “a belief is formed that is not initially prima facie justified or epistemized but which as a result of being stored in memory, becomes prima facie justified or epistemized.” (Senor 2007, 207) 2.2 Memorial Narrative Adds Doxastic Value By memorial narrative I mean a sequence of temporally ordered episodic memories with a common subject. In a narrative, reliable episodic memories can generate together new knowledge that exceeds the sum of knowledge generated by each episodic memory by itself. Memorial narrative generates “added doxastic value.” Alexandre Dumas’ novel Le Comte de Monte Cristo (1864) offers an excellent dramatic literary example: Edmond Dantès does not know why he was arrested, falsely accused, and convicted of an attempt to aid the return of Emperor Napoleon to power. After more than a decade in solitary confinement, Dantès comes to know why he was imprisoned by ordering episodic memories sequentially in a narrative with the help of’ a Socratic pedagogue, abbé Faria, who has no knowledge of Dantès’ past beyond what Dantès tells him. Reliable but individually insignificant memories lead Dantès to gain knowledge that a personal enemy with a motive to frame him witnessed events that gave him the opportunity to initiate a conspiracy to frame Dantès for a crime he did not commit. Dantès then remembers seeing later on that day the same enemy having dinner with one of his fiancée’s rejected suitors who had a different motive and further opportunity to frame him on the night prior to his arrest. Ordered sequentially, these episodic memories generate a narrative of a conspiracy to frame and denounce Dantès. In Dumas’ novel, abbé Faria contributes the empirical generalization that the handwriting of right-handed people who write with their left hand are similar and information about the Deputy who examined Dantès and his motivation for destroying evidence that could have exonerated him. I ignore this part of the plot because the rest is sufficient to make the epistemic point. The generation of knowledge through the sequencing of memories in a narrative is analogous to the plot of Michelangelo Antonioni’s film Blow-up (1966) in which a photographer takes a series of photos in a London park (analogous to episodic memories) and generates from them later, after developing them and ordering them sequentially, the knowledge that a man was assassinated in the park. When he took the photos (analogously to when the memories are formed), the photographer had no knowledge of the murder. Each photo by itself did not show an assassination. The knowledge was generated by ordering the photos as in a movie (analogously to a memorial narrative). Clearly no empirical, testimonial or introspective sources of knowledge are needed for the generation of memorial narrative. Philosophers have debated whether narrative is an a-priori basic source of knowledge--a type of reasoning, or a transcendental precondition for knowledge in the Kantian sense of the term. (Ankersmit 1983, 2008; Ball 2013) Either way, memory is a primary basic source: if narrative is transcendental, it is not a source of knowledge and cannot be said to be true or false, so memory is the only source of knowledge of memorial narratives. If narrative is an a-priori basic source of knowledge, it must be a trivial rather than primary basic source because as numerous philosophers have noted, we perceive time including our lives as temporally ordered sequence of episodes, a narrative, and we cannot experience time phenomenologically non-sequentially. (Ricoeur 1983-1985; Carr 1986) Narrative is as deeply embedded in our web of knowledge as time. If an undefeated defeater emerged to Edmond Dantès’ new beliefs about the reason for his incarceration, he would not doubt that his life has been ordered temporally sequentially, but one or more of his episodic memories. Preservationism denies that memories can generate new knowledge, claiming that if there was no knowledge at the time the memories were created, there cannot be new knowledge “out of nothing” when they are recalled. Narrative memorial knowledge proves preservationism to be false. When episodic memories are created, e.g. the episodes in Dantès’ life, the added doxastic value that emerges in memorial narratives is absent. Recall Senor’s criteria for defeating preservationism: “a belief is formed that is not initially prima facie justified or epistemized but which as a result of being stored in memory, becomes prima facie justified or epistemized.” (Senor 2007, 207) The memorial knowledge that emerges in narrative form, e.g. Dantès’ knowledge of the conspiracy against him, meets these criteria. Memorial narratives are quite common: common forms of psychotherapy generate new knowledge through ordering old episodic memories in a narrative. The writing of an autobiography may be useful for learning new things about personal history by having to order episodic memories in a narrative as well as by colligating them. 2.3 Colligation Colligation (from Latin colligere, to bring things together, a concept first introduced by William Whewell in 1840 to refer to the binding together of facts by concepts), like narrative, is a structure of reliable episodic memories as primary basic sources of knowledge through which they generate new knowledge. (Tucker 2004, 137-138) Colligation relates and synthesizes episodic memories into a single new unit of knowledge as parts into a directing whole or a pattern that is greater than the sum of its parts. For example, a senior philosopher may recall the development of her thought and identify a year that was the “pivotal year” when she broke with earlier modes of thought and sowed the intellectual seeds for her later ground-breaking philosophical contributions. During that pivotal year, she was not aware of a dramatic turn (in the Germanic sense of die Kehre) in her thought because like many original thinkers she first formulated her revolutionary ideas in terms of old and established ideas and doctrines. Colligating episodic memories in a new concept of “pivotal year” allows the generation of this new knowledge, much like historians use colligation to refer to “the Renaissance,” though nobody who lived through it thought of that period using this concept; the “Renaissance” was invented in the 19th century by the art historian Jacob Burckhardt. (cf. Walsh 1960, 59-64; McCullagh 2008) Colligation allows the generation of new knowledge from unique and reliable memories as primary basic sources. Memorial colligation is distinct from inferential memory, which is consistent with preservationism. In inferential memory, the rememberer learns a concept after a memory is formed and then applies it to the memory. The classical example is of seeing a bird, remembering it, and years later identifying which bird species it was. (Bernecker 2010, 94-96) Colligation does not involve learning and applying new concepts top-down, but colligating them from episodic memories. When the philosopher in the example above colligates that she had a pivotal year in her intellectual development, she does not learn and apply the concept of pivotal year, but colligates it from her memories. The epistemic interpretations of colligation, relating memories as parts to a greater whole or pattern, may correspond to those of memorial narrative. Colligation may be a transcendental precondition for knowledge, if it is impossible for us to perceive and remember without putting perceptions and memories into holistic patterns that are greater than the sum of their parts, or a trivial basic a-priori source of knowledge. Accordingly, colligated memories are either the only basic source of knowledge or the primary basic source. Either way, memorial colligation satisfies the Senor criteria above because the beliefs colligation generates are new and do not preserve earlier beliefs. Preservationists may attempt to preserve preservationism by dropping the Senor criteria to argue that since the memorial basic sources of narrative and colligation preserve information generated earlier by other basic sources of knowledge, they are reducible to these basic non-memorial sources. For example, though Edmond Dantès came to know about the conspiracy against him from a memorial narrative and though the senior philosopher came to know about her pivotal year from colligating her memories, the memories preserved reliably information that originated respectively from the senses and from self-knowledge. Arguably, the primary basic sources are the sources of the memories and not the memories themselves. I thank Sanford Goldberg for presenting me with this argument in our conversations after I presented a version of this paper at the European Epistemology Network Meeting in Paris in the summer of 2016. Still, dropping the Senor criteria is as counterintuitive as arguing that Edmond Dantès came to know about the conspiracy against him on empirical grounds, a decade after observing events that meant nothing to him at the time, or that the self-knowledge of thoughts the would-be senior philosopher had in her youth is the source of her knowledge of her pivotal year decades later. Surely, preservationists would have at the very least to change the meaning of preservationism to empty it of doxastic content to fit such cases. Still, there is an even stronger case against preservationism: narrative and colligation add “doxastic value” to memories that preserved information reliably. But memory can generate knowledge from no-knowledge, from unreliable memories that do not constitute knowledge. These cases should give the coup-de-grace to preservationism. 3.1 Bayesian Inference from Unreliable Memories It is possible to generate knowledge from unreliable memories that have no doxastic value. Laplace (1840, 136-156) was probably the first to demonstrate in his discoveries in the epistemology of testimony that it is possible to generate knowledge from unreliable basic sources if and only if they are coherentnowledge even from highly , independent of each other, and the prior probability of the knowledge they generate is sufficiently low. C. I. Lewis (1946, 338-362) applied Laplace’s insight to the epistemology of memory. Lewis suggested that the inference of knowledge from memories resembles its inference from testimonies à la Laplace. A single unreliable memory, like a single testimony by an unreliable witness, cannot generate knowledge. But if two or more unreliable but independent memories or testimonies convey coherent information that was considered previously improbable, it is unlikely that their agreement is coincidental. They must share a common source. (cf. Tucker 2016) The posterior probability of knowledge generated primarily from unreliable memories can far exceed the low reliabilities of individual memories if the memories are independent of each other and the prior probability of the knowledge is low. Ceteris paribus, the more information rich, e.g., detailed, are independent and coherent memories, the lower is the prior probability of the information they transmit and hence the higher is their posterior probability. This creates a psychological association between degree of detail of memory and reliability. However, the reliabilities of individual memories or testimonies are independent of their richness of detail. Fantasized memories and fabricated testimonies (like the email you received from Nigeria about a financial transaction…) can be very detailed. The degree of detail decreases the prior probability and hence is significant if and only if there are multiple coherent and independent memories or testimonies. Experimental psychologists attempt to understand how “metamemory” monitors the reliabilities of memories through “source monitoring.” When we reconstruct memories, we check for “tags” in the traces about their origins to authenticate them. The rules that regulate this unconscious process that accompanies memory construction are fallible but should be statistically reliable. The degree of detail seems to be one of these “tags.” (Michaelian 2011, 329 following Mitchell & Johnson 2000) Such tags assigned to individual memories, as distinct from multiple, independent, and coherent memories are not reliable evaluators of reliability. Here, epistemology should be “denaturalized” to normatively distinguish rational from actual-psychological evaluation of reliability, as behavioral economics demonstrated the gap between rational and actual economic behavior. To borrow Norman Malcolm’s classic example of memories of childhood kidnapping, suppose the frequency of children being kidnapped in the social milieu of David Balfour’s childhood, its prior probability, was one in a million. Yet, the adult David has clear, distinct, and independent memories of various aspects of being kidnapped as a young child. Each of these memories by itself has a low reliability because of the low reliability of memories of early childhood. But if these memories are independent of each other and coherent, the probability that they emerged in the mind spontaneously, without a common origin, is negligible. To take a quantified example, suppose Rene wonders which number won in a fair lottery in which each number from 1 to 100 had an equal chance to win forty years ago. Rene has three independent memories of the same winning number in different contexts. Each of these memories was “filed away” separately and was associated with different memories: one memory is of witnessing the draw, another of later telling a lover which number won, and yet another of boarding a bus with the same number as the winning one and noticing the “lucky number” similarity. Suppose that due to the passage of time and the insignificance of the topic, the reliability of each individual memory is a low 0.1. If we plug these numbers into the Bayesian algorithm, Pr(H|E&B)=[Pr(E|H&B)xPr(H|B)]:Pr(E|B) Pr stands for the probability of... H stands for any hypothesis, which in our case means what the memories assert, the winning number in the lottery. E stands for evidence, which in our case means the coherent and independent memories. B stands for background knowledge that assists in determining the variables (including ranges of possible probabilities). The vertical line | should be read as "given," for example, Pr(H|E&B) means the probability of a hypothesis given evidence and background information. Pr(H|B) is the prior probability of a particular hypothesis given background knowledge, for example, the odds of a winning a lottery. Pr(E|H&B) expresses the likelihood of the evidence given the hypothesis in question in conjunction with background knowledge, the reliability of the memories in this context. For example, given the winning number in the lottery, how likely it is that Rene remembers it forty years later? Pr(E|B) expresses the expectedness, the probability of the evidence given background information, whether or not the hypothesis is true. Another way of putting it formally is: Pr(E|B)=[Pr(E|H&B) x Pr(H|B)] + [Pr(E|-H&B) x Pr(-H|B)]. The posterior probability of the hypothesis given the evidence and background information, Pr(H|E&B), is the ratio of the likelihood of the evidence (coherent independent memories) given the hypothesis and its prior probability, to the expectedness of the evidence whether or not the hypothesis is true. Imagine all the possible worlds where the memories in question occur, and then ask in which fraction of these worlds the hypothesis is the case. the likelihood of the memories multiplied by the prior is 0.13 x 0.01 = 0.000001. The expectedness is 0.000001 + [(1:99x0.9)3 x 0.01]. The posterior probability that the winning number was the one in the three independent memories is then close to 0.99. The reason for this leap from individual reliabilities of 0.1 to the generation of knowledge with a close to 0.99 posterior probability is the incredibly low likelihood that all three independent memories would not preserve the information from the original event but still report the same number from among 99 possible false numbers. The lower are the priors (the more numbers could have won the lottery), the higher is the posterior probability. Bayesian inference allows unreliable memories to generate knowledge as primary basic sources. As Bovens and Hartmann (2003, 117-119) show, higher or lower reliabilities only moderate or increase the effect of the prior probabilities on posterior probabilities. Bovens and Hartmann (2003, 115-123) applied Bayesian probability to the epistemology of testimony, but it is just as applicable to any independent epistemic sources such as memories. They modelled the effects of prior probabilities on posterior probabilities in a Cartesian space, given multiple independent epistemic sources: The curve is U shaped. As the prior probabilities increase, the posterior probabilities decrease, up to a point where high priors begin to determine the posterior probabilities that rise in tandem with them. The more unreliable are the memories the deeper is the “valley” in the U-shaped curve between the high posterior probabilities that result from very low and very high prior probabilities. When the memories are unreliable, the prior probabilities have to be very low or very high to generate high posterior probabilities. For example, the prior probabilities of being kidnapped and attending kindergarten during childhood are respectively very low and very high. Therefore, the posterior probabilities of independent coherent memories of such events, even when memories are unreliable, are very high. The prior probability of meeting the president as a child is higher than being kidnapped and lower than attending kindergarten. Therefore, when the memories are unreliable, its posterior probability is lower than that of being kidnapped or of attending kindergarten. If we increase the reliability of the memories, for example by “moving” the remembered events to adulthood, ceteris paribus, the effects of the prior probabilities on the posterior probabilities moderate. Basic empirical and a-priori sources such as the frequency of child kidnapping and the Bayesian algorithm are necessary for the Bayesian inference from unreliable memories. But they are trivial basic sources. Bayesian inferences are common in science. If Bayesian inferences are more than trivial basic sources of knowledge, science is more than primarily empirical. If a defeater emerges for David Balfour’s belief that he was kidnapped as a child, he would neither doubt the Bayesian algorithm, nor the statistical data about the historical frequency of child kidnapping, but his memories of the kidnapping, the primary basic sources. Bayesian generation of knowledge from unreliable memories constitutes a decisive refutation of preservationism and reductionism because memorial knowledge that is based on unreliable memories that do not constitute beliefs let alone knowledge has nothing to preserve and nothing to be reduced into. 3.2 The Independence Problem Coherence between memories and low prior probabilities are necessary but insufficient for the generation of knowledge. (Olsson 2005, 34-35; cf. the similar criticism of C. I. Lewis in Olsson & Shogenji 2004) Coherent memories must be independent of each other. The independence of memories can be interpreted as conditional screening (memories are independent if they do not affect each other’s probability); or as causal screening (memories are independent if they do not cause each other); or as the absence of information flows between memories. (Cf. Bovens and Hartmann 2003; Olsson 2005; Tucker 2016). Memories are often conditional on, or caused by, other memories in associative chains without undermining their epistemic independence. If David Balfour has a memory of an aspect of the kidnapping that then prompts a chain of associated memories of other aspects of the kidnapping, the relations between the memories on the associative chain are causal and ceteris paribus conditional, but they would still be considered independent of each other in an epistemic sense; they can generate knowledge of the kidnapping together. Epistemically independent memories are independent of each other if the information flows they receive do not originate in other memories. Memories that receive information from other memories are dependent on them and are useless epistemically because they repeat rather than add information. It is difficult to test and justify claims for the epistemic independence of memories, for the absence of transmission of information between memories in the same mind, because there is no observable mental “ordered partition” that separates memories from each other. On the contrary, Connectionist theories of memory model it as a pattern of distributed information network. Different memories can overlap and become blended, superposition on each other. (Sutton 1998; Robins 2016, 3004-3006) The mind keeps processing the information that results in retrieved memories, arranging, rearranging, conflating, connecting, imposing coherence, consolidating and reconsolidating. (Michaelian 2011 following Ambrogi et al. 1999; Squire & Zola-Morgan, 1991; Schacter & Addis 2007) It is comparatively easier to prove the independence of testimonial basic sources of knowledge because communication between witnesses may be blocked by observable barriers. Memories cannot be wilfully stored in the mind in isolated compartments as the police can put witnesses in separate cells until they testify, nor can the mind order memories to remain separate as a judge can order witnesses not to discuss the case among themselves. When memories cohere, it may be difficult to know whether the coherence reflects receiving similar transmissions of information from common origins, past events, or exchange of information between memories. Disassociated memories may have the strongest prima facie case for independence. 3.3 Confounders Independent and coherent memories may preserve information transmitted from confounding common sources. Independent memories may generate at time t3 beliefs about events at time t1 though the common information source of these memories was an event that took place between t1 and t3, at t2. Four types of confounding sources of independent memories are possible: first, confounders at t2 may appear to be memories of events at t1, though there was no information transmission from the events at t1 to the confounders at t2. For example, suggestive therapy sessions may confound memories about early childhood without preserving any information from childhood. Repeated suggestions are most effective in generating such “false memories.” (Zaragoza & Mitchell 1996) Second, the confounder at t2 may retransmit information transmitted from t1, but misrepresent it as memories of t1. For example, somebody may form memories of a historical event from reading a book about it rather than from retaining memories from experiencing the event. The truth of the memories at t3 depends on whether the book, read at t2, was well researched. Either way, the primary basic source of the beliefs at t3 is testimonial rather than memorial. Michaelian (2016, 127-148) names such cases “contamination” or “incorporation.” He concludes that they can and often do increase the reliability of beliefs they generate and that there is nothing extraordinary about contamination because when information is retrieved from memory it often incorporates information from other basic sources of information that may outweigh the memorial source. I agree that often knowledge has multiple basic sources and sometimes testimonial sources are more reliable than memorial sources. Yet, when a testimonial source trumps a memorial source, clearly the result is testimonial and not memorial knowledge. Third, the memories at t3 may be based on information signals from both t1 and t2. For example, in oral history testimonies to memories of the slaying of an Italian demonstrator in 1949 were confounded by memories of a similar slaying in 1953. (Kenyon 2015) This is a case of memories “superimposed” on each other, as when two similar but different photographs are superimposed on each other to generate a third photograph. The resulting memory is misleading in including elements from two or more periods, some genuine and others confounding; whether the result amounts to knowledge and to what extent it is testimonial or memorial depends on the particulars or each case. Fourth, in addition to the above third type of confounding, there may be an additional information stream from t1 to t2 to t3. For example, one may form beliefs at t3 about the physical geography of the neighbourhood where one grew up both from one’s childhood memories that originated at t1 and from confounding memories of a later visit to the neighbourhood as an adult at t2. The urban neighbourhood at t2 preserved some of its features at t1, so information also flowed from t1 to t2. Again, it is impossible to generalize about whether the result amounts to knowledge and to what extent it is empirical or memorial. Tracing back the information flows from memories to the past may separate authentic memories from confounders. However, tracing information flows within the mind may have to rely on other unreliable memories or on epistemic introspection. Professionals who infer knowledge from memories like jurists, detectives, and historians use simple “folk epistemological” methods to overcome the challenges of epistemic dependence and confounders: Ceteris paribus, it is possible to lower the chances of memorial dependence and confounders by limiting the use of memories as primary basic epistemic sources to memories of recent events. The less time passes between the event at t1 that purportedly generated the memory and the time of recall at t3, the less likely it is that there was sufficient time for confounding at t2 or for the mind to process stored information flows by violating their independence. Ranke’s scientific historiography relies on the memories of eye witnesses recorded immediately after the fact and avoids relying on memoires written years after the events. (Gooch 1959, 97) The police likewise attempts to collect testimonies to memories of events immediately after they were witnessed. It is possible to discover confounders by recalling confounding events. For example, when the police interview witnesses about their memories they inquire if they discussed the events they remember with anyone previously, if so with whom, and what were the contents of those conversations. Then, they may interview then the person who may have been a “confounder” to learn what information the confounder transmitted. Oral historians ask their witnesses to try and remember if they watched movies or read books about the events they testify about. 3.4 Independent Testimonies to Memories Perhaps the most common method for overcoming the problems of memorial dependence and confounding is the use of trivial testimonies to access independent memories in different minds. It is easier to ascertain that testimonies were independent of each other than memories in the same mind were independent of each other, if witnesses could not, or demonstrably did not, communicate because of social, temporal, or geographical barriers to the transmission of information. Such barriers can be observed or inferred from independent evidential sources. It may be possible to trace back the information transmitting processes extending backward from each testimony “genealogically.” (Jardine 2008, 170-171) The reliability of information transmitted by a testimony to memory is simply the reliability of the testimony multiplied by the reliability of the memory. Since the reliability of any testimony is less than a unity, the reliability of testimony to memory is necessarily lower than the reliability of the memory itself. Therefore, ceteris paribus, unless we have reasons to consider the memories of others more reliable than our own memories, we trust our memories more than the testimonies of others to their memories. But, as discussed above, even unreliable but coherent and independent sources can generate together knowledge far more probable than any source by itself, if the prior probability of that knowledge is sufficiently low or high. We make an epistemic bargain when we generate knowledge from testimonies to memories in different minds: in return for source independence, which can generate higher posterior probabilities, we accept ceteris paribus lower reliabilities of the sources. This is an epistemically profitable bargain if exchanging reliability for independence leads to higher doxastic returns. Using testimonies to memories in different minds as sources leads obviously to asking whether the primary basic epistemic source is still memory, or testimony. The answer depends case by case on the reliabilities of the testimonies and the memories. When the reliability of testimonies to memories is very high e.g. because the testifiers have demonstrable superior integrity and no known interest in deception, the testimonies are trivial basic sources of knowledge while the memories are its primary basic source. If the reliabilities of the testimonies are more questionable and the reliabilities of the memories are high, for example if the memories are to recent events by competent adults with a long criminal record or an interest in lying, the testimonies are the primary basic source of knowledge and memory is but a trivial source. If the reliabilities are similar, both memorial and testimonial sources are equally primary. The best method for distinguishing primary from trivial sources, as ever, is to assume a defeater to the generated knowledge and intuit which source would be the first to come under epistemic suspicion. If we would examine memory first, it is the primary basic source; if we would examine testimony first, it is the primary source; if we would examine both, then they are equally primary basic sources. There are sufficient number of cases when memory is primary and testimony trivial to defeat preservationism and reductionism as I outlined above (3.1). Conclusion For much of the history of epistemology, the epistemologies of empirical and a-priori knowledge have been privileged. They have received far more attention than the epistemologies of the other three sources of knowledge: testimony, introspection, and memory. Some philosophers have defended this orthodox discriminatory inequality by claiming that testimonial and memorial forms of knowledge are not sui generis but reducible to empirical or a-priori sources of knowledge because testimonies only transmit and memories only preserve knowledge that is generated by perception and reason. Knowledge according to this view is analogous to coins that once minted by the senses or reason can only be passed from hand to hand in testimonies, or deposited in and withdrawn from memory banks, or lost. This article showed how memories as primary basic sources can and do generate unpreserved new knowledge, irreducible to the other basic sources of knowledge. Reliable memories can be primary basic sources of new knowledge in a memorial narrative, or colligated to generate new knowledge that is greater than the sum of knowledge each memory transmits by itself. Unreliable memories can generate knowledge if they are coherent and independent and the prior probability of the knowledge they generate is sufficiently low or high. The particular problems in generating knowledge from unreliable memories in one mind are in ascertaining their independence from each other and in reducing the risk of confounders. These problems may be mitigated by bootstrapping memory to recall confounding, limiting the use of memories as primary basic sources to memories of recent events, and by using reliable testimonies to memories in different minds as primary basic sources of knowledge. Narrative, colligation, and Bayesian generation of knowledge all satisfy Senor’s (2007) criteria for the falsification of preservationism, as they generate new knowledge that was not preserved in any shape or form prior to its generation from memories. I demonstrated that memory can be an irreducible primary source of knowledge in common and simple everyday contexts. Preservationists are faced now with the task of somehow explaining the range of examples I discussed here, or give up preservationism and join us, radical generativists. Sources Ambrogi Lorenzini, C. G., Baldi, E., Bucherelli, C, Sacchetti, B., & Tassoni, G. (1999) “Neural Topography and Chronology of Memory Consolidation: A Review of Functional Inactivation Findings,” Neurobiology of Learning and Memory, 71, 1–18. Ankersmit, F. R. (1983) Narrative Logic: A Semantic Analysis of the Historian’s Language. The Hague: Martinus Nijhoff. Ankersmit, F. R. (2008) “Narrative and Interpretation,” in Aviezer Tucker ed., Companion to the Philosophy of History and Historiography, Oxford: Blackwell, 199-208. Audi, Robert (1997) “The Place of Testimony in the Fabric of Knowledge and Justification,” American Philosophical Quarterly, 34, 405-422. Audi, Robert (2002) “The Sources of Knowledge,” in The Oxford Handbook of Epistemology, P Moser ed., New York: Oxford University Press, 71-94. Ball, Karyn (2013) “Hayden White’s Hope, or the Politics of Prefiguration,” in Robert Doran ed., Philosophy of History after Hayden White, London: Bloomsbury, 89-108. Bernecker, Sven (2010) Memory: A Philosophical Study. New York: Oxford University Press. Bovens Luc & Hartmann Stephan (2003) Bayesian Epistemology. Oxford: Oxford University Press. Carr, David (1986) Time, Narrative and History. Bloomington IN: University of Indiana Press. Dretske, Fred I. (1981) Knowledge and the Flow of Information. Cambridge MA: MIT Press. Dumas, Alexandre (1864) Le Comte de Monte Cristo. Paris: M. Lévy frères. Dummett, Michael (1994) “Testimony and Memory,” in Bimal Krishna Matilal & Arindam Chakrabarti eds., Knowing from Words: Western And Indian Philosophical Analysis of Understanding and Testimony, Dordrecht: Kluwer, 251-272. Gooch, George Peabody (1959) History and Historians in the Nineteenth Century. Boston: Beacon Press. Grundmann, Thomas (2011) “Defeasibility Theory,” in The Routledge Companion to Epistemology, Eds., Sven Bernecker & Duncan Pritchard, New York: Routledge, 156-166. Jardine, Nick (2008) “Explanatory Genealogies and Historical Testimony,” Episteme, 5, 160-179. Kenyon, Tim (2016) “Oral History and the Epistemology of Testimony,” Social Epistemology, 30, 45-66. Lackey, Jennifer (2007) “Why Memory Really Is a Generative Epistemic Source: A Reply to Senor,” Philosophy and Phenomenological Research, Vol. 74, 209-219. Lackey, Jennifer (2008) Learning from Words: Testimony as a Source of Knowledge, Oxford: Oxford University Press. Lackey, Jennifer (2011) “Testimonial Knowledge,” in The Routledge Companion to Epistemology, Eds., Sven Bernecker & Duncan Pritchard, New York: Routledge, 316-325. Laplace, Pierre-Simon (1840) Essai philosophique sur les probabilités, 6th Eds., Paris: Bachelier. Lewis, C. I. (1946) An Analysis of Knowledge and Valuation, La Salle IL: Open Court. McCullagh, C. Behan (2008) “Colligation” in Aviezer Tucker ed., Companion to the Philosophy of History and Historiography, Oxford: Blackwell, 152-161. Michaelian, Kourken (2011) “Generative Memory,” Philosophical Psychology, 24, 323-342. Michaelian, Kourken (2016) Mental Time Travel: Episodic Memory and Our Knowledge of the personal Past. Cambridge MA: MIT Press. Mitchell, K. J., & Johnson, M. K. (2000) “Source monitoring: Attributing mental experiences,” in E. Tulving & F. I. M. Craik (Eds.), Oxford handbook of memory, Oxford: Oxford University Press, 175–195. Olsson, Erik J. (2005) Against Coherence: Truth, Probability, and Justification, Oxford: Oxford University Press. Olsson, E. J. & Shogenji T. (2004) “Can We Trust our Memories? C. I. Lewis’s Coherence Argument,” Synthese, 142, 21-41. Plantinga, Alvin (1993) Warrant and Proper Function, Oxford: Oxford University Press. Ricoeur, Paul (1983-1985) Temps et récit, Vols I-III, Paris: Seuil. Robins, Sarah (2016) “Representing the Past: memory traces and the causal theory of memory,” Philosophical Studies, 173, 2993-3013. Schacter, D. L., & Addis, D. R. (2007) “The Cognitive Neuroscience of Constructive Memory: Remembering the Past and Imagining the Future,” Philosophical Transactions of the Royal Society of London B, 362, 773–786. Senor, Thomas D. (2007) “Preserving Preservationism: A Reply to Lackey,” Philosophy and Phenomenological Research, Vol. 74, 199-208. Squire, L. R., & Zola-Morgan, S. (1991) “The Medial Temporal Lobe Memory System,” Science, 253, 1380–1386. Strawson, P. F., (1994) “Knowing from Words,” in Bimal Krishna Matilal & Arindam Chakrabarti eds., Knowing from Words: Western And Indian Philosophical Analysis of Understanding and Testimony, Dordrecht: Kluwer, 23-28. Sutton, John (1998) Philosophy and Memory Traces: Descartes to Connectionism. Cambridge; Cambridge University Press. Tucker, Aviezer (2004) Our Knowledge of the Past: A Philosophy of Historiography. New York: Cambridge University Press. Tucker, Aviezer (2016) “The Generation of Knowledge from Multiple Testimonies,” Social Epistemology, 30, 251-272. Walsh, W. H. (1960) Philosophy of History: An Introduction, New York: Harper. Williamson, Timothy (2007) “On Being Justified in One's Head,” in M. Timmons, J. Greco, and A. R. Mele, eds., Rationality and the Good: Critical Essays on the Ethics and Epistemology of Robert Audi, Oxford: Oxford University Press, 106–22. Zaragoza, M. S., & Mitchell, K. J. (1996) “Repeated Exposure to Suggestion and the Creation of False Memories,” Psychological Science, 7, 294–300. PAGE 28