Pop culture vs measured reality · July 2026
Her imagined an AI that knew your whole digital life. It still had to find the file first.
In Spike Jonze's Her, Samantha boots up and, in seconds, has read Theodore's email, sorted thousands of messages down to the few worth keeping, and later gathers the best of his letters and ships them to a publisher. The film sells the warmth: the voice, the company. It quietly skips the part that would have to happen first: something has to actually find and relate every file on the machine, instantly, before any assistant can "just know" it. That step has a name (retrieval), and unlike the girlfriend in your ear, it is real today. We measured it on a 4.47-million-file machine.
To be clear: we didn't build Samantha. Nobody has. We built the boring, load-bearing thing beneath her, and that part is not fiction.
The 20-second version
Fiction skipped the boring part. We built it, and clocked it.
Retrieval is the first, unglamorous step of any AI assistant: finding and relating the files a request refers to, before it can reason about them. Interlinked is that layer (no model in the box, offline, private), and here is what it measures out to.
The dream is an assistant that instantly knows your whole digital life. The prerequisite the movie cut is finding and relating every file first: retrieval, plus a knowledge vault. That prerequisite is the real, measurable part.
A whole-machine file lookup returns in 85 microseconds (best case 2). That is 412,000× faster than VS Code's 35-second Ctrl+P: thirty-five seconds of a spinner versus a blink you would miss.
The content query that takes ripgrep (the tool your agent already reaches for) 93.8 seconds comes back from the warm index in 16 milliseconds: a minute and a half of scanning collapsed to less than a frame of film.
We watched Claude Code spend 6 minutes 57 seconds and 71 tool calls to find one file. The indexed version did it in one call, and stopped burning roughly 58% of the session's tokens (half its budget) on the hunt.
The vault is the “knows why it matters” half: on a fresh machine it replayed 8,455 git events across 12 repos in 1.7 seconds into 26,958 relationships. No LLM, every edge a receipt you can check.
The part the movie cut
The fantasy sits on a step the film never shows.
Her is a lovely film and a terrible spec sheet. The warm voice on top is still fiction. But it can only exist if the layer beneath it is instant and complete: find any file, relate the files that move together, recall what is already known. That layer is the whole product. Drawn honestly, it looks like this.
Same lookup · same machine · same drive
Before Samantha can know it, something has to find it.
On screen it is instant and invisible. In real life, finding one file is the slow part, the thing your agent spends its minutes and your money on. Here is how long that takes everywhere it happens today, on one machine.
| The searcher | The task | Time | vs Interlinked |
|---|---|---|---|
| Interlinked | one warm, whole-machine lookup | 85 μs | baseline |
| VS Code Ctrl+P | same file, same drive | 35 s | 412,000× |
| Windows Search | median of 20 queries | 67 s | ~480,000× |
| ripgrep (your agent's default) | one content query on the drive | 93.8 s | 58,625× * |
| Claude Code · full run | 71 tool calls to find one file | 6 m 57 s | ~7,200,000× † |
Methodology: Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4,470,000 files, July 2026. File lookups are compared to 85 μs (139 μs is the median across 20 queries; the Windows row uses it). * ripgrep's 58,625× is the average across a five-query content set; on one representative query it spent 93.8 s to what the warm index answered in 16 ms. † the Claude Code figure is the whole measured task: one call replacing the entire 6 m 57 s hunt. Full per-tool method in the 412,000× VS Code write-up and the ripgrep comparison.
Why one index isn't enough
Every code index knows one room. Samantha knew the whole house.
Cursor builds an index of the repository you have open (embeddings of that codebase), and it is genuinely good inside those walls. But it is one repo. It does not reach the other eleven that ship alongside it, the config that was never in git, the downloads folder, or the PDF with the answer. Theodore's "whole digital life" was exactly the part a single-project index cannot see.
So agents fall back to the floor: ripgrep, walking the disk on every single query, because nothing warmer exists machine-wide. Interlinked is not a replacement for the index in your editor. Keep it. It is the layer across all of them: every repo, every document, the git-ignored files a cwd search misses, handed to whatever agent you already use.
More on that boundary in how Cursor indexes your codebase and Interlinked vs the Cursor index. The whole-stack view lives in the backbone of the LLM economy.
The other half of the magic
Samantha didn't just find the letters. She knew which ones mattered.
That is the second thing the movie made look free: not just retrieving files, but relating them. The same index that answers in microseconds also carries a knowledge vault, relationships it derives by itself: files that change together, repos that share hundreds of identical files, which project is alive and which went dormant, plus short notes any agent can leave on any file, keyed to the file's OS identity so they survive a rename. No model, no embeddings: every edge is a count or a hash you can check.
How the vault seeds itself is in files that find each other. Why keying it to the file's OS identity is the whole trick is in your agent's memory dies on rename.
FRESH ON SAVE: A CHANGED FILE IS SEARCHABLE IN ABOUT A MILLISECOND, UNDER 30 MS AT WORST. IDLE, THE WHOLE ENGINE SITS NEAR 44 MB.
Questions people actually ask
The honest FAQ.
Is Interlinked the AI from Her?
No, and we like that answer. There is no voice, no companion, no model in the box. Interlinked is the unglamorous layer underneath any assistant: the retrieval and knowledge vault that finds and relates your files so the AI you already use (Claude Code, Cursor, Copilot) can reason over them. Offline, private, no keys.
What does “retrieval” actually mean here?
Retrieval is finding and relating the files a request refers to, before the model reasons about them. Measured on a 4.47M-file machine: 85 microseconds for a whole-machine filename lookup, and 16 milliseconds for a content query that takes ripgrep 93.8 seconds.
How is this different from Cursor's codebase index?
Cursor indexes the one repository you have open. Interlinked spans the whole machine (every repo, the git-ignored files, your documents) and hands results to whatever agent you use. It is parallel to your editor's index, not a replacement: keep both.
Does the knowledge vault use an LLM?
No. Every relationship is a receipt: counts, content hashes, git events you can check. On a fresh machine it seeded 26,958 relationships from 8,455 git events across 12 repos in 1.7 seconds, with no embeddings and no guesses.
What does it cost, and how do I try it?
Everything local is free, forever: the app, whole-machine search, MCP. Hosting starts at $5.99/mo. One signed installer wires up 19 AI clients at once.
Do this today
Give your agent the part the movie skipped.
Install once. One signed installer wires up 19 AI clients: Claude Code, Cursor, Copilot, Codex, Windsurf, Zed, Cline and the rest. No JSON editing.
Let it index. Filenames answer immediately; content fills in behind. At rest the whole engine sits near 44 MB, less than one browser tab.
Ask across everything. Whole-machine when you don't know where something lives; scoped to a folder when you do: either way, microseconds.
Let the vault grow. Day one it reads your git history. Every week after, it knows your machine better, on its own.
Everything local is free forever, on all your devices, no card. Hosting starts at $5.99/mo. Download for Windows. The warm voice is still fiction; the layer under it ships now.