Here is Freud recounting his memory of an incident in his infancy:
I see a rectangular, rather steeply sloping piece of meadow-land, green and thickly grown; in the green there are a great number of yellow flowers – evidently common dandelions. At the top end of the meadow there is a cottage, and in front of the cottage door two women are chatting busily, a peasant woman with a handkerchief on her head and a children's nurse. Three children are playing in the grass. One of them is myself (between the age of two and three); the two others are my boy cousin, who is a year older than me, and his sister, who is almost exactly the same age as I am. (Emphasis added.)
Freud believed this to be an accurate memory. Yet, how could it be? It is obviously impossible for someone to observe himself across a piece of meadow.
And another thing: How did he identify himself? How did he know which of the two little boys in his image was himself?
“Episodic memory” contains source information in the form of a recalled image in which the subject herself is present. (You remember what a Venn diagram looks like: you probably don’t remember learning this. Thus, you don't figure in your remembered image of a Venn diagram. This memory is “semantic”, not episodic.) The interchangeability of the two forms shows that episodic memory is not simply a stored image. Rather, it is an image that is reconstructed from elements stored in the brain. In the reconstructed image, the subject (Freud, in the above case) is tagged rather than just visually represented.
Strategies of memory reconstruction are the subject of a (rather boringly written, but) fascinating review by Daniel Schacter, Scott Guerin, and Peggy St. Jacques in the October issue of Trends in Cognitive Sciences.
Memorial reconstruction appears to be hierarchical, as the following experiment (by Guerin) shows. Subjects were presented with three pictures, which were the memory target. They were then tested by being shown images of a number of items, and asked whether these test images contained something they had seen before. When presented with an item similar to (but slightly different from) one they had seen before, they often wrongly identified the item as seen previously (as in columns 1 and 3 below. That’s no surprise, but here’s what’s odd: when presented with a previously seen item along with the subtly different item, they were able to pick which one they had seen. (The bar is the error rate: the short bar in the middle reflects the most accurate recollection.)
The errors seem to show that items are coded “semantically”, or by type. However, the middle bar seems to demonstrate that subjects also have access to the stored image.
In their review, Schacter et al use errors of memory to tease out normal memory strategies. Interestingly, they report that amnesic patients show a lower rate of certain characteristic error patterns that arise from the use of these strategies. These amnesic patients haven't just lost storage capacity: they have lost their ability to use effective memory strategies. Certain sorts of memory distortions are “adaptive”, the authors say, in the sense of being the result of efficient cognitive strategies. (This is slightly misstated: the distortions are not adaptive; rather they arise from the correct use of an adaptive strategy.)
(Rueful and largely irrelevant note: Schacter et al attribute their notion of “adaptive” distortion to Dennett and McKay 2009. It corresponds exactly to my notion of “normal error” in an article, “Teleology, Error, and the Human Immune System,” that I published in the Journal of Philosophy in 1984! It also has obvious connections to the work of Gerd Gigerenzer.)