Episodic memories enable powerful algorithms for goal-directed decision making that are context-aware and explainable

Current AI tools are often criticized for their lack of content-awareness and explainability. This make their planning and action selection sometimes undesirable for a given context, or even unsafe. In contrast, our brains can adjust planning and action selection on the fly to the current context, and often make their action selection transparent by recalling related experiences. This capability results from a fundamental difference between the way how brains and current AI tools store learnt knowledge. Whereas the latter compress all experiences into parameter values, the brain also stores representations of the most salient experiences in the form of episodic memories, which support more sophisticated action selection and planning. We show that experimental data on neural codes for episodic memories in area CA1 of the brain support a method for reproducing these capabilities in artificial devices. The resulting Episodic Neighbor(EN)-algorithm complements reinforcement learning approaches by adding inherent flexibility, context-awareness, and explainability. In addition, the EN-algorithm only needs local rules for synaptic plasticity with binary weights in shallow neural networks, and can therefore be implemented through on-chip learning in energy-efficient neuromorphic hardware.