Member-only story
Using Memories For Safety.
An Opportunity for Deliberate and Cautious Development?
How can memory be used to enhance the safety of AI agents?
First we have to look at how the structure of the memories is built and how to enable arbitrary unanticipated queries of the memories and what kinds of information is most important to store and present to users.
Some aspects of memory representations might be designed or discovered during a learning process but it should be possible to influence these representations to make safety relevant features of the remembered episodes easily retrievable.
Much more research will be needed to determine how to use episodic memories as components in an agent’s planning process.
The whole process of how episodic memories are represented, stored and accessed by an AI agent is a very important field to be aware of.
Determining what kind of training and guidance will be required for agents to learn to use the memories in a useful way:
1. How it is possible to make use of examples of undesirable actions in order to prevent similar actions in the future.
2.How to design a procedure that allows for certain kinds of newly acquired memories to influence…
