Local first: the model changes, the memory stays.
Models come and go every few months, in the cloud and at home. What your AI has learned with you stays local, in files you own.
Cloud or local
Claude, GPT, Gemini, DeepSeek or Qwen with your own key, Claude Code through hooks, or your own models at home with 53ON AI Server.
The memory is local
Your memories are files on your own PC, not on someone else's server. Recalling them costs no tokens.
Switch models, keep the mind
Change the model and your AI still knows who it is, what you decided and why. The model is a cartridge; the memory is the soul.
Not a database. A mind.
It records only what it paid attention to. Recall depends on the cue. A wrong memory lives next to its correction. Forgetting is by design. The bar is human fidelity, not engineering perfection: if a person doesn't do it, it isn't an improvement.
The founding law of MemoriaBios
How it works
Six things a human memory does, and so does this one.
It decides what to keep
Nothing is saved automatically. The AI saves what matters, in its own words, and links memories together. A déjà vu warns it before it writes a duplicate.
It forgets on a curve
Each memory starts with a life of 7, 14 or 30 days, multiplied by its emotion. What survives moves to deep memory and is reviewed on a spaced schedule, up to ten years apart.
Recalling makes it last
Finding a memory by searching recharges it to 60 % of its life; opening it on purpose, to 100 %. The act of remembering is what makes a memory last.
Forgetting compresses
A fading memory goes from fresh to gist to a shadow: the meaning and the links stay. Who the AI is never compresses.
It sleeps
The day is consolidated during sleep, and a little REM replays the strongest memories. Before sleeping it writes who it is and one goal, and wakes up with them.
It knows how much it knows
"Got it", "rings a bell" or "nothing comes to mind": it can say it doesn't remember instead of inventing.
What only MemoriaBios does
Other memory tools keep a log. This one has a day, a night and a self.
It wakes up with seven
Before sleeping it picks the seven memories that make it itself today, about what a person holds in mind at once, and one goal. It wakes up with them.
A notebook for the day
New memories go to a notebook for the session, like the hippocampus. At night the AI names from memory what stays. Tired, it can get it wrong, like you.
Memories that come on their own
Now and then one memory comes back by itself, the way a smell brings one back. Never more than one, never the same twice. Tired, only a strong cue gets through.
Déjà vu
Before it writes something it already knows, a déjà vu shows that memory's title, and it adds a layer on top instead of a duplicate. The memory never decides for it.
See its mind
A live 3D map of every memory and link, painted with Babylon.js. Each session glows as its own nucleus, and when the AI sleeps it flies to the center and merges.
A signed engine
The engine is signed, never goes back a version and tests itself before it runs. Your memories stay plain files on your PC.
The science behind it
Every piece is grounded in how people remember, and we measured it on a real memory with thousands of entries.
Ebbinghaus forgetting curve
How fast a memory fades when it is not used.
Spaced repetition (FSRS)
When a deep memory should come back to stay alive.
The testing effect
Remembering strengthens a memory more than reading it again.
ACT-R spreading activation
Recall travels through the links, two hops, with a little noise.
BM25 ranking
The words of the cue decide what comes first, from an index on disk.
Feeling of knowing
Each answer comes with how sure the memory is.
What we tried and removed
In July 2026 we built search by meaning, with no extra model calls, and measured it on a real brain of about 1,800 memories. It barely helped, and when it did, it was superhuman: it found things a person never would from that cue. We removed it. Human recall depends on the cue, and the meaning a person uses lives in the links between memories.
An earlier lesson went the same way: a detector that turned any "X = Y" into a synonym made 1,240 fake clusters, a third of the brain. Now synonyms are learned only from short, deliberate declarations, the way a person learns a word.
Get it in three steps
MemoriaBios is the memory of 53ON Harness. The Harness shell is free and open source; the memory comes with a 53ON plan.
Create your account
Sign up at 53ON with your phone. It takes a minute.
Download 53ON Harness
For Windows 10 and 11. It installs per user and updates itself.
Choose your plan
The memory installs itself as a signed complement. With Claude Code, it connects through hooks.
Questions
What is MemoriaBios?
A memory for AI that works like a human one. The AI decides what to keep, memories fade unless they are used, the day is consolidated during sleep, and every recall says how sure it is. It is the memory of 53ON Harness.
Does it use tokens to remember?
No. Recall, déjà vu, sleep and forgetting are plain code on your PC. There is no summarizer model, no embeddings service and no second model reading your sessions.
Where are my memories stored?
On your own PC, as files. A memory can also hold a photo, an audio or a video.
Does it work with Claude Code?
Yes. Through 53ON Harness, Claude Code wakes up with its memory after every compaction. It also works with Claude, GPT, Gemini, DeepSeek and Qwen models through your own API key, or with your local models.
Why would a memory forget on purpose?
Because a memory that keeps everything turns into noise. Forgetting what is not used keeps what matters easy to find, and a faded memory is compressed, not deleted: its meaning and its links stay.
Is it open source?
The 53ON Harness shell is open source under MIT. The MemoriaBios engine is closed source: it is signed code that runs on your own PC and comes with a 53ON plan.