The vault series · 3 of 5
Not RAG. Not memory. The layer under both.
There are two famous ways to give an AI something like memory: a vector database it retrieves from (RAG), and the model's own saved recollections (ChatGPT-style memory). They feel like opposites, but they share a spine: both live on the model's side of the line, both answer approximately, and neither can see the actual files on your machine. The vault is a third thing. It doesn't compete with either; it runs underneath both. Exact, deterministic, no model in the loop.
Two approximations, one truth
Two layers of guessing. One layer of truth.
RAG turns your files into vectors and stores them in a database like Pinecone; at query time the model pulls back the nearest matches by meaning, approximately. Model memory is the feature that quietly saves summaries of what you tell it and threads them into later chats. Useful, both of them. But each is a model-side sketch of your world, and neither one can read a file on your disk. Read the diagram top to bottom: the sketches on top, the world itself underneath.
The two upper layers are the model's approximations: one indexed by meaning, one by what you happened to type. The layer beneath them is the world as it actually is: exact file identity, literal contents, and relationships derived from real history, every one a receipt you can check. RAG and model memory guess at what the vault simply knows.
Three ways to remember, side by side
Two of them guess. One of them checks.
| Stores | Keyed to | Freshness | Its answer is | Scope | |
|---|---|---|---|---|---|
RAG / vector DB Pinecone-style | vectors (embeddings) | a vector, not a file | re-embed to update | a ranked guess | what you ingested |
Model memory ChatGPT-style | summaries you typed | your vendor account | when the vendor saves | a paraphrase | your conversations |
The vault Interlinked Files | identity + literal text + edges | the file's OS identity | ~1 ms after a save | a receipt you can check | every repo + non-git files |
None of these are bad tools, and they aren't rivals: keep your vector database for meaning, keep your model's memory for continuity. The gap is structural: a vector is keyed to an embedding and a memory is keyed to an account, so neither can be pinned to a file or verified against one. Only a machine-wide index can key knowledge to a file's own OS identity (the thing that survives a rename), and that is what lets it be exact instead of approximate.
The reason it's a different layer
Both live above the OS. The vault lives at it.
A vector store answers from what you ingested and embedded. Add a file, rename a function, refactor a module, and the vectors keep pointing at the old meaning until you re-embed. Retrieval is nearest-neighbor (approximate by design), so it's superb for discovery and shaky at “did I catch every exact call site?” It never touches the file; it touches a snapshot of it.
Model memory stores what you said in a conversation, keyed to your vendor account. It never saw your files, so it can't tell you which one changed this morning, which repo shares code with which, or where a symbol actually lives. It's a paraphrase of your intent, held by someone else's server. Genuinely useful, and structurally blind to the machine.
The vault is the only one of the three with OS-level access: it indexes the whole machine, so it can key every note and every relationship to a file's permanent OS identity, watch that file so a save is searchable in about a millisecond, and reach every repo plus the roughly 90% of the disk that was never in git. RAG and model memory can't reach down there: they sit a layer up, in the model's world. That access is the entire difference between a guess and a receipt.
The thesis
Keep both upper layers. Put the exact one underneath.
RAG answers “what is this probably about?” Model memory answers “what did I tell you?” The vault answers a question neither can: “what is actually on this machine, right now, exactly?” It is the ground truth the other two approximate, so it belongs in the same stack, running parallel to both and sitting under both. No model, no embeddings, no vendor holding your recollections: every edge is a fact the operating system already recorded, turned into a count you can verify by hand. Keyed to file identity, it survives renames, spans every repo, and never leaves the machine.
Methodology: Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4,470,000 files. Seed figures are from a fresh install, timed from the service log before the first query. The same index answers a content query in about 7 to 9 ms (up to ~16 ms for an exact phrase) and sits near 44 MB at rest. The exact-retrieval and knowledge layers are the one engine.
THE VAULT SERIES CONTINUES: THE MACHINE-TRUTH LAYER, NOTE BY NOTE.
Put it under your stack
Keep RAG. Keep memory. Add the truth layer.
Nothing here asks you to drop your vector database or turn off your model's memory. They answer questions the vault doesn't. What the vault adds is the layer neither can be: exact, deterministic knowledge keyed to your files, fresh to the last save, spanning every repo plus the 90% of the disk that was never in git, and read by every agent you run. One signed installer auto-configures 19 AI clients (Claude Code, Cursor, Codex, Windsurf, Zed and the rest) with no JSON to edit. At rest the whole engine sits near 44 MB, and the on-disk index is under 1% of the drive.
Everything local is free forever, on all your devices, no card. Hosting starts at $5.99/mo. Download for Windows · Previously: files that find each other