zygon4/mmesh

theoretical AGI memory system

Theoretical AGI memory system as a mesh network. (sort of - we're still in flux)

It's a twist on the HTM CLA. It's sort of an ant swarm intelligence concept. The hope is that it's more scalable and simpler. It's really just a fun thought exercise for now.

The basic idea:

• You have a network of cells
• Each cell listens for activations from remote cells
• If that same active cell was predicted to be active by a different cell, then it sends a prediction feedback message and that prediction becomes stronger
• Right now, the activations and predictions are time-expiring (for better or worse - this is still in flux)

The hope:

• Anomaly detection
• Scalable learning (small local tables leading to a distributed memory mesh, no need for billions of synapses (read: pointers))
• Online learning

What isn't yet:

• There's no conversion from actual outside data to SDR which is presumed to be fed into the mmesh as input activations