VaultHelm sits between the AI you already use and the things you actually care about. Your AI carries on doing what it’s good at. Everything it does gets written down, and anything that matters has to come past you first.
It’s just files. That’s on purpose.
Everything VaultHelm keeps is stored as ordinary text documents. Not a database, not a format only we can read. You could open any of it in Notepad tomorrow.
Right now we host that for you — your own separate copy, with its own files and its own key. It is never a shared system with several customers’ work in one pile. And because it is only a folder of text files, you can take a copy of the whole thing, full history included, whenever you ask.
Underneath, every version of every file is kept, so nothing is ever really deleted — you can always go back and see what something looked like last month.
This is deliberate, and it is the answer to the obvious question about letting someone else hold your record. If VaultHelm vanished tomorrow you would still have a folder of text files that opens on any computer, with its full history intact. A record you cannot read without the company that sold it to you was never really your record.
One record, every AI you use
This is the part worth being clear about.
Every assistant keeps its own memory, and none of them share. What you told one is invisible to the next. Each vendor is quietly building a silo, and the longer you use one the more expensive it becomes to leave.
VaultHelm sits underneath all of them instead. Your record and your permissions live in one place that belongs to you, and every assistant you connect reads and writes that same record. Ask any of them what happened on Tuesday and you get the same answer.
Be clear about what does and does not move. Nothing is piped between AI companies. Your conversation with one is never handed to another, and we are not brokering your data between vendors. What carries over is the written record on your side, and only what you chose to file into it. Each assistant reads that record the same way it would read any document you gave it.
In practice that means:
- Several at once, not one at a time. This is the part people assume is switching, and it isn’t. Have three of them open together on the same job — one drafting, one checking the first one’s answer against the record, one on your phone in the van. Each reads the same filed record when it starts, so none of them begins cold and none needs briefing.
- Approvals travel as well as the record. A request raised by one AI can be approved from a different one entirely. Proposed at your desk in the morning, approved from your phone that evening, in another app.
- A better model costs you nothing. When something new launches you just try it. Everything you’ve built up carries over, because it was never stored inside the AI to begin with.
Grok, Claude and ChatGPT all connect today, natively — inside each one’s own app, on desktop and on phone, with nothing extra to install. New assistants can connect to the same record without you rebuilding that record inside them, and when an AI company changes how its assistant plugs into outside tools, re-doing the connection is our problem, not yours — on your side it is a setting you turn on once.
The honest caveat about running several at once
They do not see each other’s work the instant it lands. When one of them files something there is a short catch-up before the next one can read it. For nearly everybody that gap is invisible, because you are not usually asking two assistants about the same sentence in the same minute — you are asking one of them about last March.
If you want that gap closer to nothing, it can be tuned, and it costs more to run. More frequent checking means more work done on your behalf, and that is a real bill rather than a setting. We would rather price it honestly than pretend the delay isn’t there, and we will tell you what the gap actually is on your setup before you are ever asked to care about it.
Reading at the same time is free and unlimited — have as many open as you like. Writing is the part that is deliberately arranged, so that two assistants working at once cannot quietly overwrite each other. You never have to coordinate them, and you never discover a fortnight later that one of them stamped on the other’s work. That is the same principle as everything else here: the guarantee is in how it is put together, not in asking the assistants to be polite to one another.
One word at the end
When you finish working, you type one word. That’s the whole ritual. VaultHelm then:
- Keeps the conversation exactly as it happened, word for word.
- Sorts it out — if you worked on three different customers today, it splits it into three.
- Files each piece under the right customer’s history.
- Closes off whatever job you were working on.
- Tells you what it did, including how much of your work it accounted for.
That last one matters more than it sounds. Summaries lose things. So before VaultHelm tells you it saved successfully, it checks how much of your actual work made it in. If too much went missing, it keeps the whole conversation word for word instead of a summary. It isn’t allowed to say “done” without doing that check.
And if your connection drops halfway through, just say the word again. Repeating yourself never creates a duplicate — it only fills in what didn’t make it the first time.
Finding it again
Weeks or months later, you ask a normal question — “what did we do for the Henderson job in March?” — and you get three things:
- The entry. The short version: who it was for, when, what it was about.
- The write-up. What was actually done, in that customer’s own history.
- The original. The raw conversation it came from, untouched.
You can always get to that third one. If the summary got something wrong you can correct it, and the correction sticks — tell it once where that work belongs and it files there from then on. But the original conversation underneath is never rewritten. Corrections change the index, never the record.
What you can ask it that you can’t ask a chat window
The useful questions aren’t “summarise my week”. They’re the ones that need two things compared:
- Which source wins. The live record from the supplier and the spreadsheet from 2022 disagree. You get told there’s a conflict and which one is current, rather than a confident blend of both.
- What’s still open. Take what was left hanging after an incident, check it against everything filed since, and say which items nobody ever went back to. Nothing filed is a real answer here, and it’s usually the valuable one.
- How old this is. An answer that comes off a write-up from three years ago says so, so you don’t spend the morning working from a drawing that stopped being true.
- What was quoted, and what was worked out. There’s a difference between “the file says this” and “I put this together from two files”, and you get told which one you’re looking at.
It will also tell you when it couldn’t look — a file it isn’t allowed to open, a source that didn’t answer. A refusal you can see beats a confident answer built on something that was never read.