M-Logic Machine General Structure of Memories

When I talk about sensory memory, short-term and long-term memory, and many other terms found in psychology and cognitive sciences, like memory decay and interference, I am really just giving names to the machinery and the semantics of the M-Logic Machine.

The general memory structure of the M-logic Machine can be seen in Figure 1. The names given are justified by the obvious similarities with the multi-store model of human memory. However, the cinematic structure of memories and other MLM assumptions brings some important differences. The distinction between episodic and semantic memories becomes fuzzy, since the same structure can provide both types of information. The transfer from short to long-term memories is not made by rehearsal, and no single store is assumed for each type of memory.

Figure 1. The General Memory Structure of the M-Logic Machine

I call sensory memory (rather a binding memory) a short cinematic memory (just two frames) where the fullness of the present-moment recordable sensory data is kept permanently available for further selection and recording. This is needed because the short-term memory will have to ignore lots of data, trying to retain the most essential features in a given context for a longer time. This means the short-term memory becomes blind to many aspects of the sensory data. Each filtering is called a "sensory mode". But some vital information may need to call back the attention of the agent and change the current selection. The sensory memory keeps the machine alert. The instruments that feed the sensory memory are called cover-set intruments. A cover-set instrument maps some measured feature to distinct positions in space (in other words, distinct output lines are used for distinct results).

The short-term memory is permanently recording from the sensory memory some selected data, keeping ten to twenty frames of the recent past. This interface between sensory and long-term memories is most needed because recording in the long-term memories is intermittent. Some heuristics are used to wisely start and stop recording from the short-term memory. For instance, one obvious criterion to start recording is "pain", the data brought in by the nociceptors of the agent. In order to avoid in the future a similar painful situation, the machine may consult its long-term memories. But this would be pointless if the available cinematic data only started from the moment pain aroused. The machine therefore needs to find in its short-term records what happened just before the recording started. This vital information about the recent past is offered by the short-term memory, and inserted in the long-term memories when recording starts.

Long-term memories need to be of several kinds. For the purpose of long-term inference that uses several scenes, a chronological order of the recordings of actual data must be kept in a real long-term memory. For the purpose of reinforcement learning in a setting of short-term predictions (i.e. using a single scene), a dominance mechanism that selects the most accurate cinematic memories is implemented in a dominance-list memory. Internally generated scenes and tales (for instance, those resulting from some kind of planning) are kept separated in an imaginary long-term memory. Long-term memories are compartmented according to the sensory modes used, leading to something I call "compartmented learning".

When heuristics are used to provide inference, data is retrieved from sensory, short-term, and long-term memories. This retrieved information is centralized in a working memory. The configuration of data in the working memory - depending on what could and could not be retrieved - defines the epistemic states of the machine.