Why graphs beat vectors for agent memory

RAG chunks your documents, embeds them, and returns the top-k most similar at query time. It's great for lookup and terrible for reasoning — because similarity throws away structure. Grep has the same problem in a different costume: it finds strings, not relationships.

The multi-hop wall

Ask 'what breaks if I change this function's signature?' and a vector store returns chunks that mention the function, or signatures, or breakage — and hopes the model stitches them together. Grep finds every file containing the name, including comments, tests, and an unrelated variable that happens to shadow it. A graph follows the edges: function → callers → their callers → the schema they serialize. That's a traversal, not a guess.

Structure is extracted, not embedded

Graphify builds those edges from the AST — tree-sitter parsing, done locally, across 36 languages. A call edge exists because the parser found the call, not because two chunks landed near each other in embedding space. Where the model does help — connecting docs, SQL, and Terraform to the code — the resulting edges are tagged INFERRED so you can tell them apart from what was extracted.

A score isn't an answer

The quiet failure of similarity search is that it can't explain itself: a cosine score of 0.83 tells you nothing about why a chunk is relevant. A traversal is its own explanation — the answer is the path, and every node on it is a real file you can open. That's the difference between retrieval that sounds right and retrieval you can check.