Regarding modelling memory: Years ago, I was just for fun implementing different strategies for NPCs to walk around and interact with each other (little dots an a hex map). The NPCs had different personalities. For example one was an explorer prefering to move to unknown areas on the map. Another was staying at home and foraging. They could than exchange information of the world (memory) for food. Here the memory was quite specific: Each NPC had his own map of the world consisting of his personal knowledge and what he aquired by information. Some memories were static (the feature of the terrain), others dynamic (the last know location of another particular NPC). I had no concept of fading or inacurate memory. In the case of dynamic memories, the latest information was just overriding the older.

> Theoretically, how would you imagine separating memories? Would you use memories clusters with different weight for example ?

There are, of course, a lot of possibilities. I was always fascinated by the idea to use something like a Semantic Web as the basic structure to represent memory, because it is so flexible. Every atomic memory element is a subject-predicate-object or subject-predicate-value triple. We can build bigger memory structures just by repeatedly using the same entities for subjects or objects. The predicates may be predefined ontologies and/or themselves be represented as (higher order) subjects/objects modelled by very fundamentel predicates (like "<A> is a subtype of <B>"). We could model the fading of memory by removing atomic memory elements from the set. We could model the blurring of memory by slightly changing an atomic memory element (like "<Bob> has the haircolour <brown>" into "<Bob> has the haircolour <dark-brown>". For decision making, we could use an algorithm that explores the set memory elements, draws conclusion, prioritieses them according to some evalution function. The possibilites here are endless.

In this model, the combination of certain predicates, the structures resulting from them and of particular algorithms that operate on them are constituting a memory cluster. For example, if the NPC is hungry, this may activate an algorithm that looks for the predicate "<x> is of type of <Food>", then for every x found looks for "<x[1]> is located at <y>", then for every location tries to determine a path to that location by connecting memory elements ("The cheese is in the fridge", "The fridge is in the kitchen", "The kitchen is adjacent to the living room", "The armchair is in the living room", "The NPC sits in the armchair").

Going back to my old example of getting food offered I dislike when I am very hungry, this could be modelled as: "<Offer> is of type <Event>", "<Entity-4758> is of type <Offer>", "<Entity-4758> has actor <PC>", "<Entity-4758> has receiver <NPC-6587>", "<Entity-4758> has transfer-item <Item-8974>", "transfer-item <Item-8974> is of type <Apple>", "<NPC-6587> dislikes <Apple>", "<Entity-4758> has point in time <T-7125>", "<NPC-6587> has physical state <P-7894>", "<P-7894> has point in time <T-7125>", "<P-7894> is of type <Hunger>", "<P-7894> has intensity '0.95'" (rang from 0..1).[1]

To simulate remembering the situation, we can than run an evaluation function on this data set that gives us a valence/arousal vector. This function may have other parameters (for example the current mood of the NPC) that can modify the outcome of the evaluation. Or we can simulate the fading of memory by randomly modifying the value of predicate "<P-7894> has intensity '0.95'" towards neutral, or the object of "transfer-item <Item-8974> is of type <Apple>" into <Banana>. Both of which would change a later evaluation of event <Entity-4758>.

Not all of the predicates from my example represent memory in the strict sense. Some represent built in ontologies that should not change ("<Offer> is of type <Event>") or treated as immutable facts to keep things simple ("<NPC-6587> dislikes <Apple>"). Our algorithms that modify memory must not change them (unless we want to simulate that a NPC starts to become crazy).