Maybe I missed it in the paper but they did post the source code (Github) for their implementation? Is anyone working on creating their own infrastructure based on the paper?
Their approach to memory is interesting. I had been considering a tiered command-based approach -- "short-term memory" being an automatic summary of recent sensory inputs/command outputs; "long-term memory" being a detailed database queryable by the agent.
Yes I plan to make it open source.