Agents · opinion, with data · July 2026
Every AI startup is secretly a search company.
Pull the chat box off the hottest AI products and the same machine hums underneath every one: a search engine. Perplexity searches the web. Glean searches your company. Cursor searches your repo. Even “memory” is a lookup. The model is the part everyone shows off; the retrieval is the part that decides whether the answer is any good. This is the case that most AI products are retrieval problems wearing a model, and where the last missing search layer lives.
Strip the chat interface off almost any AI product and what is left is a search engine.
RAG is search. Agent context is search. Model memory is a lookup. The interface hides it; retrieval does the work.
The breakout AI companies are retrieval companies: Perplexity indexes the web, Glean indexes the enterprise, Cursor indexes the repo.
One retrieval layer is missing: the local machine, for the agents on it. That gap falls back to grep.
Interlinked is that layer: a whole-machine index that turns a 93.8-second content scan into 16 milliseconds, with no model in the box.
The mask comes off
Five products. One engine underneath.
Retrieval is the step where software finds the right information and puts it in front of the model before it answers, ranked, filtered, and handed over. It is search. Read the diagram left to right: how each product is sold, and what it is once the interface is peeled away. The pattern is hard to unsee.
The last row is the one nobody built. Perplexity, Glean and Cursor each stood up a real index for their patch of the world. The AI agent on your own machine got none, so it falls back to grep, re-reading the disk on every question. That empty box is the whole opportunity. It is also the backbone of the LLM economy, the layer every agent stands on, and the one nobody optimized.
The features you already use
Five AI features that are search in a trench coat.
None of these are fake AI. The model is real and useful. The point is narrower and sturdier: in each one, the thing that decides whether it works is a retrieval step, and the retrieval step is a search problem someone has to solve well.
The name literally starts with retrieval. A query runs against an index, the top hits come back, the model reads them. The retrieval decides whether the answer is any good; the generation is the easy part.
Before an agent edits a line it has to find the file, the function, the call site. Every one of those is a search, run over and over, and mostly with grep.
“Remembering” is pulling the right saved fact into the prompt at the right moment. That is ranked retrieval over your history, dressed up as recall.
Glean and its peers do not hide it: the product is search across everything a company owns, with a chat box on the front and permissions on the side.
Perplexity calls itself an answer engine. Under the answer is a web search and a model to phrase the citations: the search is the substance.
The through-line: the model is a commodity you rent; the retrieval is the thing you own and the thing that differentiates you. That is why “memory” is a feature war right now: it is a retrieval war wearing a friendlier word.
Two the market already crowned
Perplexity and Glean never pretended to be anything but search.
ANCHOR 1: THE WEB
Perplexity calls itself an answer engine.
Its own public word for the product is not “chatbot”: it is answer engine. That is the tell. Under the tidy paragraph and the footnoted links sits a web search, and a language model whose job is to phrase and cite what the search found. Take the retrieval away and there is no product; take the phrasing away and you have a worse Google. The retrieval is the substance, the model is the manners.
This is not a knock: it is the whole reason it works. A better answer engine is a better search engine first. The famous name is the doorway; the durable business is retrieval.
ANCHOR 2: THE ENTERPRISE
Glean is enterprise search that grew a chat box.
Glean was built by search engineers to index everything inside a company (the docs, the tickets, the chats, the wikis) with permissions baked in, and only later put an assistant on top. The moat was never the model; it is the permission-aware index of the whole company. The assistant is what you see. The index is what you buy.
Web and enterprise are two of the three big retrieval surfaces, and each has a champion. The third is the one you are reading this on: the local machine, with a swarm of agents already running on it and no index built for them.
The same X-ray, in a table
What each one is really doing.
Read the “where it stops” column top to bottom. Every champion indexes one surface and stops at its edge. The last row is the only one whose surface is the machine itself.
| The AI product | Sold as | The retrieval job underneath | What it indexes | Where it stops |
|---|---|---|---|---|
Perplexity answer engine | ask anything | rank the web, cite the links | the public web | can't see your files |
Glean work assistant | your company's AI | search across company apps | your SaaS + docs | stops at the company cloud |
Cursor · Copilot coding AI | knows your code | index + retrieve the repo | the repo you opened | stops at the workspace |
ChatGPT memory assistant memory | remembers you | look up saved facts | what you told it | blind to your disk |
A RAG app custom assistant | trained on your data | vector search over docs | what you ingested | only what was embedded |
Interlinked local retrieval layer | search for your agents | index the whole machine | every repo + non-git files | on your machine, offline |
“Sold as” is each product's own public positioning; the retrieval column is what the feature does mechanically. None of these are rivals of each other: they cover different surfaces. The gap they share is structural: none can index the files on your local disk for the agents running there. More on the retrieval spectrum in RAG vs index vs grep.
The surface nobody indexed
The web has Google. The enterprise has Glean. Your machine has grep.
There are three retrieval surfaces that matter to an agent: the open web, the company's cloud, and the local machine it actually runs on. The first two got purpose-built indexes worth billions. The third (the one with your code, your configs, your documents, and the dozen repos that ship together) never did. When an agent needs a file there, it shells out to the same primitives a 2006 laptop used: walk the directory, grep the bytes, hope.
That is not a small tax. On a real 4.47-million-file machine, one whole-machine content query took 93.8 seconds with ripgrep, a minute and a half of the machine reading every file, on a loop, every time the agent asks. The fix is not a smarter model. It is the missing index.
The third champion
We built the index for the machine.
Interlinked Files is the local retrieval layer: a purpose-built, always-fresh index of every file name and every file's contents across the whole machine, every repo, including the git-ignored files a working-directory grep never sees. A filename comes back in 85 microseconds; the content query ripgrep spent 93.8 seconds on comes back in 16 milliseconds, a minute and a half of reading every file versus a single frame of video. It rides the filesystem's own change signal, so a save is searchable in about a millisecond, and it hands results to 19 AI clients over the one interface they already speak. No model in the box.
And the index remembers how your files relate. Derived, not guessed.
The same engine carries a knowledge vault, relationships it works out from your own history: files that change together, repos that share hundreds of identical files, which project is alive and which went dormant. On a fresh machine it read 8,455 git events across all 12 repos in 1.7 seconds and drew 26,958 relationships, and flagged two repos that share 435 byte-identical files, unprompted. Every edge is receipts you can check, never a model's guess.
Methodology: Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4,470,000 files. Filename lookups are median of repeated runs; the ripgrep row is a content query over the same corpus; the agent run is the measured Claude Code session. Full per-tool method in the 412,000× write-up and the ripgrep comparison.
THE WEB HAD ITS SEARCH COMPANY. THE ENTERPRISE HAD ITS SEARCH COMPANY. THE MACHINE HAS ONE NOW TOO.
Questions people ask
The short version, in Q&A.
Is every AI company really a search company?
Not literally, but strip the interface off most AI products and what does the work is retrieval: finding the right context to put in front of the model. The chat box is the surface; the search underneath decides whether the answer is any good. The teams that break out tend to be the ones with the best retrieval.
Isn't RAG different from search?
RAG is search. “Retrieval-augmented generation” begins with retrieval: a query runs against an index, the top results come back, and the model reads them. The generation is the last step: the retrieval decides the quality.
What is a local retrieval layer?
A local retrieval layer is a prebuilt, always-fresh index of everything on your machine (every file's name and contents across every repo, plus the roughly 90% of the disk that was never in git) that answers in microseconds and hands results to any AI agent. It is search built for the machine, not a single folder.
How is this different from Perplexity or Glean?
Perplexity indexes the public web; Glean indexes your company's cloud apps. Neither can see the files on the machine in front of you. Interlinked is the retrieval layer for the local machine and the agents running on it: offline, private, every repo included.
Does it use AI or embeddings?
No. It is an index and a match: no model, no embeddings, no keys, offline. The relationships it derives (which files change together, which repos share files) come from receipts like git history and content hashes, not a model's guess.
Give your agents the missing surface
Put a search company under your machine.
You already trust retrieval for the web and the enterprise. The local machine is the surface your agents actually run on, and it is the one still stuck on grep. Interlinked is the index that fixes it: whole-machine, always fresh, offline, no model in the box.
Install once. One signed installer auto-configures 19 AI clients: Claude Code, Cursor, Copilot, Codex, Windsurf, Zed, Cline and the rest. No JSON editing, no keys.
Let it index. Filename search works immediately; content search fills in behind it. At rest the whole engine sits around 44 MB, and the index on disk stays under 1% of your drive.
Search whole-machine, or scope. Whole-machine when you don't know where something lives; scoped to a folder when you do. Both answer in the microsecond-to-millisecond range.
Let the vault grow. Day one it knows your git history. Every week after, it knows your machine better, automatically, with no model in the loop.
Everything local is free forever, on all your devices, no card. Hosting starts at $5.99/mo. Download for Windows · The five ways an agent finds a file, ranked