The vault series · 2 of 5
Files that find each other.
You don't write the vault. It writes itself. The first time Interlinked runs on a machine, it reads the one deep history that already exists there (your git commits) and webs the whole thing together. Here's what that looked like on a real machine, timed from the service log, before a single search was run.
TL;DR
Files that find each other, in five lines.
- 01The relationship vault seeds itself the first time Interlinked runs, no setup. On a real machine it webbed 12 of 12 repos in 1.7 seconds and derived 26,958 relationships from 8,455 git events, before a single search.
- 02Every edge is a receipt, not a guess: a fact the OS already recorded (a shared commit, an identical content hash, a literal reference), turned into a count and checked against a floor. No model, no embeddings; you can verify any edge by hand.
- 03The connections ride along on searches the agent already runs: one search returned launcher-health.tsx plus a free edge to drive-map-graph.tsx (co-changed 6×), the exact file the fix needed next.
- 04It runs the 28-year-old "files that change together are coupled" law live and machine-wide (across every repo and the non-git 90% of the disk), offline, on your own machine.
- 05Git history is only the seed; your behavior is the harvest. The longer Interlinked lives on the machine, the more it knows, and it never leaves.
No model. No embeddings. No guessing.
Every connection is a receipt, not a hunch.
An embedding tells you two things are probably similar. We don't do probably. Every edge comes from a fact the operating system already recorded (a shared commit, an identical content hash, a literal reference) turned into a count, checked against a floor. You can verify any edge by hand in one step: open the git log, grep the string.
The same machinery runs on more than commits: identical content hashes across repos ("shares 435 identical files"), literal references ("named in 106 files of the other"), shared rare symbols and config values. Every note names its count, so a wrong conclusion is visibly wrong instead of silently absorbed.
Day one, in production
The graph named the next file before the agent knew it needed one.
An agent was fixing a UI control it had never seen before. It searched for the control by name and got the file it expected: launcher-health.tsx. Riding along underneath that search result, unasked, was a single line from the vault:
launcher-health.tsx ⟷ changes-with drive-map-graph.tsx · co-changed 6×That second file was exactly where the fix needed to continue. The agent didn't have to grep for it, reason about it, or get lucky: the connection was already there, derived weeks of git history ago, waiting on the search the agent was always going to run. It was right twice that day: the same edge pointed the way through a second, unrelated bug an hour later.
This is the difference between search and a vault. Search finds the file you asked for. The vault knows which file you're about to ask for next.
This idea is 28 years proven
"Files that change together are related" is old science. Running it live, machine-wide, is new.
Researchers proved in 1998 that modules changing together are coupled in reality, even when no code reference connects them. CodeScene built a respected business on it: behavioral code analysis, sold to engineering managers. But it mines one organization's repos, in batch, on a server, at enterprise prices.
Sourcegraph's code graph needs a server too, and only sees what's been pushed. Ours saw that two repos on the local disk share 435 identical files (one is the other's ancestor): unpushed, unconfigured, day one. We take the 28-year-old law and run it the way nobody else can: live over the change journal, across every repo and the non-git 90% of the disk, included in a free tool, on your machine, offline.
What happens after day one
Git history is the seed. Your behavior is the harvest.
The install-day seed is the deepest history that exists on the machine, but only for code. From then on, the vault watches what actually happens: the files you save together, the documents you open in the same sitting. That tax PDF and the spreadsheet you always open beside it become a relationship no git repo ever recorded. The longer Interlinked lives on your machine, the more it knows, automatically, and it never leaves.
NEXT: NOT RAG. NOT MEMORY. THE LAYER UNDER BOTH.
Common questions
The vault, asked and answered.
How does Interlinked connect files with no AI or embeddings?
Every edge is a receipt, not a hunch. It takes a fact the operating system already recorded (a shared git commit, an identical content hash, a literal reference), turns it into a count, and keeps it only when the count clears a floor of three. Because no model runs, any connection is verifiable by hand: open the git log, grep the string.
Do I have to configure it or feed it anything?
No. The seed runs on its own the first time Interlinked starts: every repo on the disk, webbed from its own history, in about 1.7 seconds. From then on it learns from what you actually do: the files you save together and the documents you open in the same sitting.
How does my agent actually receive these connections?
They ride along on the searches your agent was already going to run, no extra call. Search is effectively instant (a single-file lookup in about 85 µs, content search in ~7 to 9 ms), and the relevant edges come back underneath the result, so the agent learns the next file to open without grepping for it.
Does the vault stay current, and is it heavy to run?
It stays live. Edits are reflected in about 1 ms through the OS change journal, and the service idles at roughly 44 MB of memory. It works entirely offline, and the graph never leaves your machine.
What does it cost?
Everything (including the MCP layer that feeds these relationships to Claude Code, Cursor, and the rest) is free forever, no card. Hosting starts at $5.99/mo.
See it on your own machine
Install it tonight. It knows your repos by morning.
There is nothing to configure and nothing to feed it. Install Interlinked, and the seed runs on its own: every repo on the disk, webbed from its own history, in seconds. Ask your agent "what repos do I have?" and you'll get biographies, not a directory listing.
Everything local is free, forever: the app, whole-machine search, MCP. Hosting starts at $5.99/mo. Download for Windows · Previously: your agent's memory dies on rename