july 2, 2026 · 1 min read · rag, agents, memory, relay
From chunks to atoms
Agent teams are quietly moving off chunks plus top-k. What they need looks less like retrieval and more like structured state with provenance.
Six months ago every AI-native team was rebuilding their RAG pipeline. Today they're deleting them.
Not because retrieval failed. Because "retrieve chunks of a document by embedding similarity" is turning out to be the wrong shape for what agents actually need.
I've been building a memory layer for AI tools. In every conversation I have with agent team founders now, someone eventually says "yeah, but what we really need is…" and describes something that isn't retrieval at all. It's structured state with provenance.
Text chunks don't have owners. Decisions do. Chunks don't supersede each other, but policies do all the time. Chunks are stateless string blobs. A real memory primitive carries a version chain and a citation back to its source.
The whole vector-DB versus graph-DB debate feels like it's happening one level too low. The question isn't "which store", it's "what's the primitive". If the answer is still "a text chunk with an embedding", we haven't updated our model since 2023.
I could be wrong on this. Maybe RAG holds and the primitive just gets richer. But every serious agent team I've talked to in the last month has been quietly moving off the "chunks plus top-k" architecture toward something that looks a lot more like a knowledge graph of atoms.
Originally posted on LinkedIn.