Why not files

Most teams keep what they know in a folder of notes, or in a vector store of chunked documents. Both work. Both fail on the same four things, and they fail for the same reason: the text is stored, but nothing about the text is stored.

The four things text does not carry

	A file or a flat vector dump	A cortex
Identity	"Dana Ito", "[email protected]" and "D. Ito" are three unrelated strings.	Canonical normalization collapses them into one node, so a fact about any spelling is a fact about the person.
Relationships	Proximity only. Two facts are related if they happen to sit near each other in a document, or near each other in embedding space.	Typed edges. A row is from a person, owned_by an organization, attended_by a set of people, because a rule said so.
Provenance	A chunk, sometimes with a filename.	A universal_id that resolves to the exact source row, in every layer, plus a graph link that highlights it.
Ranking	One similarity score, which says how close the text is, never how trustworthy the claim is.	Every membership and edge is tagged with how it was derived, from a declared field down to a probabilistic extraction, so you can rank by kind of evidence.

Add more documents and none of those four improve. Identity gets worse: more spellings, more duplicates, more near-identical chunks competing for the same slots in a result set. Search over a pile of text returns more text. It does not return an answer.

What a graph gives you instead

1. Canonical identity

   Extraction rules normalize a value before it becomes a node: emails, phone numbers and URLs each have their own normalization. Two spellings of the same person land on one node. See typed entities.

2. Typed edges

   Rules also connect rows that share a common value at the same path, which is the mechanism that turns flat rows into a graph rather than a list.

3. A trust ladder

   Every edge records how it was produced. A declared field outranks an exact lexical match, which outranks a geometric similarity, which outranks an LLM extraction. See the derivation ladder.

4. Citations

   Every result carries universal ids and a graph link. A claim you cannot trace is a claim you can drop.

Retrieval, not search

The practical difference shows up in what your model receives. A vector dump hands it the nearest chunks of text and leaves the reading to the model. A cortex hands it a small set of scored, typed, cited results, so the model's context carries answers rather than a haystack to search.

That is also why the same setup helps a small or local model rather than only a large one. See context and inference cost.

What files are still good for

Nothing here says stop keeping files. Files are the source. A vault, a Drive folder and an export are all normal inputs, and a row can point at a file on the office drive through file_path while its text lives in the graph. The argument is about what you query, not about where the original lives.