AI Memory That Works on a Plane (No Internet, No API Key)

Somewhere over the Red Sea, my AI got amnesia

I was on a flight with no wifi, trying to finish a piece of work I had already thought through properly weeks earlier. The notes were on the laptop, right there in a folder, a few centimetres from my hands. But the assistant I was talking to could not see any of it, because the part that remembers was sitting on somebody else's server, and that server was not on the plane. So I did what everybody does. I started retyping context from memory, badly, and got a worse answer than the one I had already written down.

That flight is a big part of why Callosium works the way it does.

No key, no internet, no cloud in the core

Callosium turns a folder of plain Markdown notes on your own computer into one memory that every AI you use can read and write. The part that does the finding runs entirely on your machine. It answers with no key, no internet, and no cloud in the core. There is no large language model in the loop when it searches, and no paid API quietly running up a bill behind the scenes.

The only time it touches the network at all is to sign in and to check for updates. No note ever leaves your device to run a search. Nothing is uploaded to a service to make the search work. That is not a privacy mode you switch on. It is just how the thing is built.

So how does it find anything with no internet?

Several independent searches (I call them lanes) run at the same time, and their results get combined. Each lane is good at something the others are bad at:

• Title ladder. Catches the easy case where you more or less named the note you are asking for.
• Keyword ranking. Classic word matching, except the title, the headings, and the body each carry a different weight.
• Section coverage. Scores how much of your actual question a note's best passage answers, giving more credit to rare words. A note containing your one specific unusual term beats a note that only shares filler like "project" and "notes".
• Proximity. Rewards notes where your words sit close together instead of scattered across ten pages.
• Graph. Pulls in notes linked to the strong matches, so surrounding context rides along.
• Semantic. A small multilingual model living on your laptop (it turns text into numbers so that things which mean the same thing land near each other) finds notes that say your idea in completely different words.
• Temporal. Wakes up only when your question carries recency in it ("latest on", "what did I do last week", and the Arabic equivalents) and pushes fresh notes up.

The lanes get merged using Reciprocal Rank Fusion, which is a heavy name for a simple idea: compare who ranked what first, so the lanes never have to agree on a scoring scale. The whole core is deterministic, meaning the same question over the same notes gives you the same answer, every time, on your hardware. If you want the longer version of how to set the folder up in the first place, I wrote that in build a second brain every AI can read.

Local is also just faster

People assume offline means slower. It is the other way around. On a single machine, a typical request lands in 49 milliseconds. Nine out of ten come back inside 77 milliseconds. The slowest request in a hundred takes 116. No round trip to a data centre, no queue behind other customers, no rate limit at the worst possible moment.

The token question, answered honestly

Every time your AI reads context, someone pays for it in tokens (the chunks of text a model charges by). Callosium reads about 3,400 tokens per answered question. Here is the honest comparison, all of it:

• Against the common habit of dumping the top five matching notes in full, which costs around 255,000 tokens, Callosium reads about 99% less.
• Against a careful keyword agent reading snippets, around 5,100 tokens, Callosium reads about 35% less and answers far more often.
• Against a perfect oracle handed the single correct note, around 990 tokens, Callosium reads more. On purpose. It returns a small cited cluster instead of one lonely note, which is exactly what lets your AI verify the claim and follow the thread.

I am not going to sell you a flat "90% savings" number. The 99% figure is against the wasteful habit. Against a tidy oracle, I lose, and I would rather tell you that than have you find out later.

Why most "local" tools are not really local

Plenty of tools carry the local-first label. Then you try to install one and it wants Docker, or a database you have to run and maintain, or it quietly falls back to a paid model the first time a question gets hard. Callosium is unzip and run. No terminal. I put together a fuller comparison in the roundup of local-first AI memory tools if you want to check my work against the alternatives.

It works in Arabic, not as an afterthought

The local model is multilingual rather than English-only, so English and Arabic work at parity. That matters most for paraphrased questions, where you asked in one wording and wrote the note in another. More on that in an AI that works in Arabic.

And if you walk away tomorrow, your notes stay

Delete Callosium and your folder is still sitting there, plain Markdown, readable in any app you like. Nothing to export, nothing held hostage, no format only I can open. That is the whole point of notes that outlive your subscription.

The core is free forever and open source under Apache-2.0. Early access opens 4 August 2026, and the waitlist is on the homepage. Take it on a flight and see if it still knows you.

Abdo

If your memory needs a signal to work, it was never really sitting with you.

Frequently asked questions