The project that made it click.

Magpie-Search made something click for me.

I kept thinking of ICM and vector search/RAG as competing ideas, but they're not.

Vector search, embeddings, and RAG were never meant to be the memory—they're the retrieval layer.

ICM is the memory architecture.

What's interesting is that modern LLMs already do a small version of this. They don't keep everything in context all the time—they selectively retrieve what's relevant because context is expensive.

The difference is that most systems only do this when it's convenient.

What if that wasn't the exception? What if it was the architecture?

An ICM-first approach means memory is always organized into structured containers, and retrieval is always responsible for bringing only the relevant context forward. Instead of trying to fit everything into a giant context window, the AI continuously builds the context it actually needs.

That's the part that feels like the real game changer to me.

It isn't ICM or vector search/RAG.

It's ICM with vector search/RAG.

0 comments

The project that made it click.