Change the AI.
The conversation continues.
You can switch AIs. You cannot switch yourself. SOKI is a memory of your own that lives outside the AI. What you told ChatGPT, Claude already knows. And it does more than remember: when a date draws near it opens with a “by the way”, it brings back your “I’ll do it later” with a “come to think of it”, and before the AI sounds too sure it adds a “just to be safe” — a memory that brings things back to you. Every memory carries its date and how certain it is, and you can review and correct it yourself at any time. Set it up once; after that, just talk as you always do.
- One memory across ChatGPT and Claude
- “By the way” for dates, “come to think of it” for your laters
- A “just to be safe” before it asserts
- Every memory dated — review and fix it yourself
All you do is talk, the way you always have.
There is nothing to press to make it remember, and nothing to press to make it recall. It works quietly, inside the conversation.
Talk as you always do
Talk to ChatGPT or Claude exactly as you did before. There is no form to fill in.
SOKI remembers
It picks out only the facts from what you said. It does not write down judgements about you, such as "meticulous".
It recalls in the next conversation
At the start of a conversation the AI comes and looks at SOKI. What you told ChatGPT, Claude already knows.
This is not just a tool that remembers.
Remembering is only the entrance. What SOKI cares about is working from any AI, adding a word before it asserts, and remembering what you said you would do and never closed. None of the three exist in memory that lives inside an AI.
The same memory across every AI
Neither memories nor skills live inside one particular AI. What you told ChatGPT, Claude knows. Switch AIs and you do not have to move yourself.
Learn more →Not at face value — a word of caution
When the AI asserts something, SOKI adds a line about what that answer is standing on. It does not rewrite the answer. It only hands you something to check.
Learn more →Plans and "I will do it later", inside the conversation
Just say "car inspection on the 10th next month, tell me three days before". No calendar entry, no reminder to set. It also keeps the things you said you would do and never closed, and asks you about them quietly.
Learn more → Try it on your notes →Skills — call a favourite instruction by name
Register a way of asking that worked — "minutes", "a polite refusal" — and from then on just say "minutes, with the skill". You create them, edit them and switch them off yourself.
Learn more →Memories have weight
What changes easily fades quickly; what rarely changes stays. Each time something comes up again it grows stronger, and a plan grows stronger as its date approaches. An old memory is not treated as reliable forever.
Learn more →Your memories stay in your hands
You can see everything it remembers as a list, in the words you used, and export all of it at any time. Sensitive information such as health is not recorded by default.
What it will not remember →Do not keep your memoryinside the AI.
SOKI does not belong to any one AI. Change the AI you use and you do not have to move yourself.
If it lives inside the AI
- The day you leave that AI, everything it remembered ends with it.
- Every move to another AI starts from the introductions again.
- Use two AIs side by side and neither knows what the other heard.
If it lives outside the AI
- What you told ChatGPT, Claude knows.
- When a new AI comes along you just reconnect. The conversation continues.
- The skills you registered can be called from any AI too.
- When you leave, you can export everything and take it with you.
It works with AIs that follow the MCP standard. Step-by-step guides are provided for Claude and ChatGPT, and both have been checked on real machines.
A word of caution
on the AI answer.
An AI answers with confidence even when it does not know. When it asserts something, SOKI adds the assumption the answer stands on and one thing worth checking.
- It does not rewrite the answer. It never says "this answer is wrong". It only adds what to check.
- It checks against your own memories. If an answer contradicts a fact you told it or a decision you made, it says so.
- A decision is kept together with its reason. Later it asks once: "does that reason still hold?" So the decision does not outlive the reason for it.
- It is on from the start. Turn it off in settings if you do not want it. It stays quiet in small talk.
It is meant as a nudge, so that an answer is not taken at face value. Especially for people who have just started using AI.
What error is shown on the screen?
Got it. I will bring it up in conversation from the 7th.
No calendar.
No reminders.
What you said in a conversation comes back in a conversation. SOKI never asks you to "add an entry".
- Just say the plan. "Tell me three days before" and "in the morning" are remembered with it. As the date nears the AI mentions it in conversation, and afterwards asks how it went.
- It does not forget your "later". Things with no date — "waiting for a reply", "not decided yet" — are kept too, and asked about quietly when the conversation suits. It does not nag.
- A morning note by email. For days you do not talk to an AI, and only if you ask for it: times and counts only. The contents are not written in it. Nothing is sent on an empty day.
- No push notifications. Not being noisy was decided at the very start of the design.
- There is a limit on how often it speaks up. Reminding and pressing are not the same thing. The number of times SOKI may break into one conversation is capped, so it does not derail what you were saying.
Try it right now. No account, no waiting.
SOKI learns as you use it, so normally it takes a while of talking before things start coming back to you on their own. So that you do not have to wait, here is a page that does the same thing with the notes you have already written. An Obsidian vault, or any folder with .md files in it.
Try it on your own notesNo sign-upYour files are never uploaded. Everything runs inside your browser — it works with Wi-Fi switched off.
Call the way of asking that worked
by its name.
"Turn this into minutes." "Write a polite refusal." You do not have to get the wording right every time. Register a way of asking that worked as a skill, and from then on the name alone gives you the same result.
- Four are included from the start. Minutes, a polite refusal, translate-and-summarise, and line drawing. Use them as they are, or rewrite them.
- You can write your own. Just a name and a body in the settings screen. Up to 20 each. Edit any time, and pause the ones you do not use (nothing is deleted, so you can bring them back).
- They work across AIs. A skill registered from ChatGPT can be called from Claude. Nothing is locked to one AI.
- Only you can write them. An AI cannot register or change a skill. That is so a web page you had it read cannot slip an instruction in without you knowing.
minutes, with the skill is all you sayPlease use the SOKI skill "minutes" naming the skillThe wording differs slightly between AIs (this is what actually worked when tested). The same guidance is in your settings screen.
(pastes the meeting notes)
I have tidied up the notes.
- ■ Decided: exhibit at next month’s trade show
- ■ To do: get a quote (owner: A, no deadline set)
- ■ On hold: the logo refresh moves to next time
- 📝Minutes
- ✉️Polite refusal
- 🌐Translate and summarise
- ✏️Line drawing
- ＋Your own skill
Like human memory —
some fades, some stays.
A memory in SOKI is not simply present or absent. Each one carries a weight that moves over time. So something you said last year is not treated forever as if it were true today.
- What changes easily fades sooner; what rarely changes lasts. Something like where you were born does not fade at all.
- Each mention makes it stronger. Memories you use often last longer; ones that do not come up quietly move to the back.
- A plan works the other way: it grows stronger as the date nears. Once it has passed it asks how it went, once, and then fades.
- Say "that is no longer true" and it is corrected. It is suppressed rather than deleted, so the history can still be traced.
- A memory of middling confidence is checked before it is used. "Is this still true?" — answering is what updates it.
First, what it cannot do.
Better to know before you sign up than to find out afterwards.
What it will not remember
- Health, beliefs, social status and other sensitive personal information are not recorded by default. They are handled only if you allow it in your settings.
- People other than you are recorded under pseudonyms only, and the mapping to real names is not kept on our server.
- Only statements of fact are recorded. Judgements about a person — "meticulous", "anxious" — are not.
Yours
- You can see the whole list of what it remembers, at any time.
- You can export all of it and take it with you. Nothing is left behind when you leave.
- Delete or close your account at any time. Connected AIs can be disconnected at any time too.
- Basic records such as your name or where you are from cannot be retracted on the AI's own judgement. Correcting them is your call.
Set it up once.
You connect from an AI that supports MCP (Model Context Protocol). All of the setup is one time only; after that we never ask you to do anything.
First — SOKI cannot be used alongside the AI's own memory.
Please turn off the memory feature in the AI you use. If it stays on, the AI writes what you said into its own store, and nothing reaches SOKI. It does this silently, so it looks as though everything is working. It is not strictly required for Claude, but two stores means two places to look, so off is recommended.
Choose the AI you use
The desktop apps keep their settings in a different place from the browser versions. Pick the one you actually use.
A connected AI can be disconnected at any time.
SOKI is used by connecting it to an AI such as ChatGPT or Claude. It does nothing on its own. See how to connect above.
Before you begin.
MCP, "server" — this sounds technical. Do I need to know any of it?
No. All you do is paste the SOKI URL into your AI's settings screen. There is a guide with screenshots for each AI. After that you just talk as usual — there are no commands and no special operations.
Which AIs does it work with?
Any AI that supports MCP. Step-by-step guides are provided for Claude and ChatGPT (browser and desktop app for each). One SOKI can be connected to several AIs, so your memories carry over when you switch.
Can I use it alongside the built-in "memory" in ChatGPT or Claude?
We do not recommend it. With the AI's own memory switched on, it may write what you said into its own store and nothing reaches SOKI. Worse, it does this silently, so it looks as though everything is working. Please turn your AI's memory feature off. (It is not strictly required for Claude, but two stores means two places to look, so off is recommended.)
How do I create a skill?
Log in, open "Skills" in your settings, write a name, a description and the instruction for the AI, and save. Four are included from the start, so the quickest route is to rewrite one of those. Skills you do not use can be paused (nothing is deleted, so you can bring them back). Asking the AI to "register a skill" will not work — only you can write them.
Where are the memories stored, and who can see them?
They are held on SOKI's server. Only you can view, export or delete your own memories. Third parties are recorded under pseudonyms only, and sensitive personal information is not recorded by default. See the privacy policy for details.
Can I stop partway through?
Yes. You can disconnect any AI at any time, and export, delete or close your account whenever you like. Nothing is left stranded when you go.
What does "beta" mean here?
That it is being built in the open. Screens and behaviour may change, and some things may not work yet. If you notice something, please tell us through contact.
Never explain yourself twice.
Set it up once. After that, just talk as you always do.