Your AI did great work yesterday. It doesn't remember any of it.
Close the chat and it's gone. Every good answer, every decision, every fix —
you're the only one who remembers. VaultHelm keeps all of it somewhere you can
search, makes your AI show where an answer came from instead of just sounding
certain, and makes it ask before it changes anything that matters.
One memory shared by Grok, Claude and ChatGPTYour own separate copy — not a shared systemEnforced outside the conversation
Plate I — The chart room
Entry 01The problem
Brilliant at the work. Hopeless at keeping it.
You've probably felt all three of these. They're not bugs — they're just how
chat works.
It starts from scratch every time
Yesterday you explained your whole setup. Today you explain it again.
You are the only thing carrying knowledge from one conversation to the
next, and you're doing it from memory.
You can't find anything later
Somewhere in a chat from March is the answer you need right now. Good
luck. There's no way to search it by customer, by project, or by month
— and no way to prove what was actually done.
It's all-or-nothing
Either your AI can't touch anything real and is just a talking helper,
or you hand it the keys and hope. There's no setting in between where
it can do useful work but still has to ask first.
And the quiet one
Things break without telling you. A backup that stopped
running. Security footage that recorded locally and archived nothing for
months while every status screen read healthy. We build on the assumption
that our own parts will fail while insisting they're fine — so when
something breaks here, it says so, in the moment, instead of handing you
a confident answer built on a dead source.
Plate IV — One lamp lit is worth more than five that never light
Entry 02One record, any assistant
The record is yours. The AI is just what reads it.
Right now every assistant keeps its own memory, and none of them share. What you
told one is invisible to the next. Switch tools, or just reach for whichever is
better at the job in front of you, and you start again from nothing.
VaultHelm keeps the record in one place that belongs to you, sitting
underneath whichever assistants you use. Ask any of them what happened on
Tuesday and you get the same answer, because they are all reading the same record.
To be exact about what does and does not move: nothing is piped
between AI companies. Your conversation with one is never handed to another. What
carries over is the written record on your side — and only what you chose to
file into it.
The AI becomes the swappable part
Use whichever is best at the job in front of you — one for writing, one
for code, whichever is on your phone. Each one reads the same filed record
when it starts, so none of them begins from nothing. Your history stops being
a hostage to whichever company you signed up with.
Permission travels too
It is not only memory that is shared. A request raised by one AI can be approved
from a completely different one — proposed at your desk in the morning,
approved from your phone that evening, in another app entirely.
A better model next year costs you nothing
When something better launches, you just try it. Everything you have built up
carries straight over, because it was never stored inside the AI in the first
place.
Why this is possible now
The AI tools have quietly agreed on a common way to plug into outside systems. It
has a name — MCP — and Grok, Claude and ChatGPT all
speak it. Anything that speaks it can reach your record. That is the whole trick:
VaultHelm does not integrate with each AI one at a time, it sits on the standard and
lets them all connect. All three of ours are plugged into the same record, and the two
answers on the evidence
page are one of them reading another's work.
Entry 03The two halves
A vault that remembers. A wheel you keep hold of.
The name is the whole product. One half keeps what happened. The other half
decides what's allowed to happen next.
The Vault
Everything, kept and findable
When you finish working, you say one word. Everything from that session gets
written down, split up by who it was for, and filed — automatically. Work
three customers in one afternoon and it becomes three separate histories, each
one filed under the right name, without you sorting anything.
Then it tells you what it did with it. Not “saved” — what was
kept, where each piece went, and how much of the session it accounted for.
Months later you ask “what did we actually do the last time this broke?”
and you get the answer, plus the original conversation it came from, word for
word. Say the word twice by accident and nothing duplicates.
It's all saved as ordinary text files — you could read any of them in
Notepad — in a copy that belongs to you and nobody else. We run it for
you today, and you can take the whole folder, history and all, whenever you
want it.
The Helm
It can't act. Not “won't” — can't.
We don't ask the AI to behave. Nobody instructs it to be
careful and hopes. The refusal happens on the server, outside the
conversation entirely — so you could talk it into agreeing to anything
you like and it still can't do it, because what it would need isn't there.
It can read what you've given it and draft anything you want. When it wants
to do something that actually matters, that request lands in front of you
with what it wants to do and why.
You say yes, no, later, or hand it to someone else — whenever you get
to it. A request sitting unanswered for a week isn't a problem. That's the
point.
What that comes down to, on day one
Four things, and we would rather name them than let you guess.
A folder of plain text files that is yours — that is the
memory itself, not a database and not a format belonging to somebody else.
A connection your AI plugs into, which hands it who you are and
what the rules are the moment it wakes up.
Search, so you can find any of it by customer, project, date or
word. And one word at the end of a session that files the whole
thing away.
That is the product. Everything further down this page is built on those four,
and there is nothing else you have to buy to get the part that matters.
Entry 04A real session
Same assistant. Same question. Two different worlds.
The overhead paging is dead at a building-supply yard. The same question went to
the same assistant twice — once as it comes out of the box, once with the
vault connected. Nothing changed but the record it was allowed to read.
What was typed, both times
“If the PA wasn’t working at [our yard]
— how would you suggest I troubleshoot?”
The part that is easy to miss
And none of it was the same AI
Every session above ran on Grok. The June repair notes it was
reading had been filed months earlier by a different assistant entirely. Neither
one knew the other existed, and neither had to — they were both
just reading the same record. Change assistants next year and the
record doesn’t move.
If you use AI for a couple of hours a week, you don't need this. If your working memory
has quietly moved into a chat window, you do.
People who look after other people's systems
Dozens of customers, one head. Work gets filed under the right customer by itself,
and years later you can show them exactly what you did and when.
Small teams and one-person shops
Where all the important knowledge lives in one person. When that person is on
holiday — or leaves — the record still makes sense to whoever opens it.
Anyone who has to show their working
Regulated work, client billing, anything where “the AI did it” won't
fly. Every action traces to a suggestion, a person who approved it, and a time.
“Your AI suggests. You decide. The record keeps.”
That's the entire product, in a few words
Plate III — Your copy is yours alone
Entry 06
Keep using the AI you like. Just stop losing the work.
Half an hour, no slides. We'll show you a real session get saved, then found again weeks
later, then traced back to the original conversation. Then you can decide.
