EM-LLM: Human-Inspired Episodic Memory for Infinite Context LLMs

TTT, cannon layers, and titans seem like a stronger approach IMO.

Information needs to be compressed into latent space or it becomes computationally intractable

So, infinite context length by making it compute bound instead of memory bound. Curious how much longer this takes to run and when it makes sense to use vs RAG.