An offline AI that actually remembers you

Run a local model on its own and you get something genuinely impressive and genuinely blank. It can explain soil biology or draft a letter, and it has no idea who you are. Ask the same thing twice tomorrow and you'll reintroduce yourself twice.

That's not a bug in the models — it's how they're built. A model is a frozen piece of training plus whatever text is in front of it right now. If your life isn't in front of it, it isn't in the answer.

For a developer this is fine; they're building something and don't need a friend. For everyone else, it's the reason local AI often gets installed on a Sunday and forgotten by Wednesday. The demo is clever. The daily experience is talking to an amiable stranger.

Your Konshus Vault holds memories drawn from your conversations, imports and notes. When you link Sovereign, it copies those memories down into a private local Vault, encrypted on your computer. Resync any time to bring down what's new.

Copying runs one way only — Online to computer. Nothing flows back automatically.

The reason for that direction is simple: your Online Vault is where your memories are built and kept. Your computer is a copy you carry with you. Letting the copy quietly overwrite the original would be a way of losing things, so Sovereign doesn't do it.

• Your Konshus greets you by name and picks up recent themes.
• Ask about something in your Vault and it can tell you, with a note on which memories it used.
• When two memories disagree, it favors the newer one and says when it's unsure.
• If your Vault doesn't contain an answer, it says so rather than inventing one.
• It can tell you what it doesn't have — which turns out to be as useful as what it does.

The last two are the ones worth testing. Ask your local AI something your Vault definitely doesn't contain and watch whether it admits that. An assistant that guesses is worse than one that says it doesn't know.

The obvious fix for the stranger problem is to dump your entire chat history into the model's context and let it read. People try this. It doesn't work, for two reasons.

First, room. Every model can only hold a limited amount of text in play at once. A few years of conversations overflow that immediately, so you end up with a truncated slice — and the model doesn't know which parts were dropped.

Second, relevance. Most of a chat history is small talk, half-formed thoughts, dead ends, and things you've since changed your mind about. Feeding all of it in doesn't make an AI know you better; it makes it know you noisier.

What works instead is a Vault: memories that have been pulled out, kept current, and checked against each other, so the AI starts each conversation with a short, accurate picture rather than a wall of transcript. When it needs more, it can look, and say which entries it used.

Had a great conversation on a plane? Save it as a file from Sovereign and add it to your Online Vault when you're back. It's deliberate by design — Sovereign never sends your chats anywhere on its own.

This is the part people expect to be a sync and are glad it isn't. A background sync would mean your local conversations — the ones you kept on your computer precisely because they were private — quietly becoming server data. Saving a file yourself keeps the decision where it belongs.

It also makes the useful case easy: a trip's worth of thinking, added to your memory in one deliberate act, instead of a stream nobody chose.

Your local Vault is encrypted and protected by your Recovery Key. You can erase it completely from the app, or with the separate remover if the app won't open.

The Recovery Key cuts both ways, and it's worth being blunt: because we never store it, we can't recover your Vault if you lose it. Keep the written copy somewhere safe, the way you'd keep a spare deed or a safety-deposit box key.

One honest limit: your own backups. If you back your computer up with Time Machine or something similar, copies made there are yours to manage. An old backup copy can't be opened without the key that was deleted — but you still need to know it exists.

The AI running on your computer is a different model from the one behind Konshus Online, and it's usually a smaller one. So the style can differ a little — a local brain may be more direct, or less polished on a long complicated request, especially on a smaller machine.

What doesn't differ is who it is: the same name, the same memories, the same Vault behind it. People describe it as the same friend on a quieter day.

You can switch brains at any time without losing your conversations or your Vault, so if one brain's voice isn't clicking, another may. And the memory side keeps improving with each release, which is the part that actually decides whether it feels like yours.

You can check in five questions whether a memory setup is doing anything. Ask them in order, and read the answers as a reader would:

• "What do you know about me?" — you should get a short picture, not a shrug and not a transcript.
• "What have I been thinking about lately?" — this tests whether recent material is reaching it.
• "What do you know about my [person, place, or project]?" — this tests the specific entries you added.
• "What's something you don't know about me?" — a good assistant admits the gap instead of inventing one.
• "Which memories did you use for that?" — you should be able to point at them.

A vague answer to the first three usually means a thin Vault, not a broken brain — go add the material and resync. A confident wrong answer means a stale or mislabeled memory, and it's worth correcting at the source rather than arguing with the model.

Run the same five on a plain local model with no Vault at all and the difference is immediate: it can only tell you what you just typed. That comparison is the fastest way to understand why memory is the part worth paying attention to.

It's tempting to treat memory as a nice extra on top of a model — the feature that makes the demo friendlier. In practice it's the difference between a tool you consult and one you talk to.

A model without memory answers your question. A model with memory answers the question you were actually asking, because it knows the context you didn't bother to repeat: the dog's name, the job you're between, the trip you're planning, the fact that you already explained this last month.

That's also why the memory has to be yours. If the picture of your life lives in one company's account, it can be trimmed, reset, or stranded when that account changes or the product changes direction. A Vault you can copy down to your own machine is a memory you can keep.

The model side of local AI improves on its own. Newer brains arrive, answers get steadier, and more of them fit ordinary computers — whether or not you do anything.

The memory side only improves if the material is there. A year from now, the difference between an AI that knows you and one that merely sounds clever will be a pile of ordinary notes you added along the way: the trip you took, the decision you made and why, the thing you changed your mind about.

That's the quiet advantage of keeping your memories in a Vault. They keep working for you no matter which brain happens to be running today.

It doesn't learn from your offline conversations automatically — those stay in the conversation until you deliberately save the chat as a file and add it Online. That's a choice, not an oversight, and it's the same one-way rule that keeps your private chats from quietly becoming server data.

It doesn't know current events, and a smaller brain will occasionally be shallower or blunter than the one behind Konshus Online. Style differs between brains; the memories don't.

And it can still be wrong about you in a way that feels convincing. When that happens, don't assume the brain is broken — ask which memories it used and check the entry behind the answer. Most of the time the fix is one stale memory, not a new model.

• Add to your Online Vault as you go — notes, documents, the things you'd want your AI to know.
• Confirm memories when Konshus asks, and correct the ones that drifted. A confirmed memory is a reliable one.
• Resync before a trip, not during it, so the copy on your laptop is current.
• Ask it "what do you know about me?" and read the answer. Where it's vague is where your Vault is thin.
• When it's wrong, say so plainly. Then check the memory behind it rather than assuming the brain is broken.

A Vault is a garden, not a filing cabinet. Ten minutes of tending a month changes how well your AI knows you more than any brain upgrade will.