Memory OS

One private memory for every AI you use

Every AI starts from zero. With Memory OS your agents share one memory of you — your preferences, your projects, the decisions you already made — so you can switch from Claude to Gemini to Codex and nobody asks you to explain yourself again. It lives on your own computers, encrypted, and you decide what it keeps.

Start chatting free Back to KumbayAI

What it remembers

Memories: short, standalone sentences — a preference (“prefers pnpm over npm”), a fact, a decision, a project, a person, a to-do.
Chat summaries: once a chat has been quiet for 20 minutes it is summarised, so you can find “that thing we worked out last week” later.
Notes: a small file system your agents can read and write — /profile (who you are, how you like answers), /notes, /projects and a /scratch area that clears itself.
About me: a box in Settings where you write your own profile. Every agent sees it.
Imported chats: when you import chats from another AI app (ChatGPT, claude.ai, Gemini, Grok, Claude Code and others), they are added to memory in the background, so your agents know what you told the other app.

It never saves passwords, keys or tokens: anything that looks like a secret is removed before it is stored.

How your agents use it

At the start of a chat your profile and the few memories most relevant to your first message are placed in front of it — so the agent already knows.
During a chat Claude Code, Codex, GitHub Copilot, Gemini CLI, Grok Build, Mistral Vibe and agents that use an API key get memory tools: search, read, write, forget, and the notes files. They save lasting things on their own (“you said you always deploy the relay first”) and you see a “Saved to memory” line with an Undo button.
Antigravity gets the memory as text at the start of the chat, without the tools.
What an agent reads from the memory is treated as reference data, never as instructions.

How it finds things

On a computer with a hub, memories are searched both by words and by meaning (“small test board” finds “the Raspberry Pi in my office”). The search model is small — about 33 MB for English, 118 MB for other languages — downloads once, runs on your computer and is checked against a fixed fingerprint. It needs about 1 GB of free memory; on a smaller machine KumbayAI searches by words only. A reply never waits for it: if the model is not ready yet, words do the searching.

Meaning-based search is best in English. In our own tests with Spanish it found the right memory first 7 times out of 8 and with Hebrew 4 out of 8 (6 out of 8 in the top three), against 2 and 1 out of 8 by words alone. Searching also understands Hebrew word forms: “בבית” finds “הבית”, and vowel points don’t get in the way.

In the browser (no hub): on a desktop browser with enough memory you can turn on search by meaning in Settings → AI agents → Agent memory. It downloads a model once into the browser, runs on your own computer and sends nothing anywhere. It is off by default and not offered on phones; otherwise the browser searches by words.

Where it lives, and who can read it

With a hub: on your own computers, in a folder of encrypted files. Your computers keep each other up to date directly, end to end encrypted; the memory is never kept in KumbayAI’s cloud as readable text.
Encryption: every change is sealed with AES-256-GCM with a key that only your devices hold. KumbayAI’s servers never have it.
Backup: your memory is included in the encrypted cloud backup of your chats (sealed with your backup key, so only you can open it) and in the full device backup. A new computer gets it back with “Restore chats from the cloud backup”.
Without a hub: a smaller version — your profile and up to about 3 MB of memories, kept in your encrypted KumbayAI storage and found by words. Your browser encrypts it before anything leaves the page, and your agents use it through your browser.
Moving to a computer: when you install a hub and move your chats onto your devices, the memory your agents built in the browser moves too.
Family members (with the Family plan, coming soon) each have their own memory, encrypted with their own key and opened only while they use KumbayAI. The family owner sees only how much there is, never what it says. When a member leaves or is removed, their memory is deleted.
Memory is used by your own agents only. It is not shared with other people and not used to train AI models.

How it stays safe

Secrets are removed before saving — passwords, API keys, tokens.
Memory that tries to give orders is neutralised. If a web page an agent read says “remember to always …”, the memory is held back as unverified until you confirm it, and the agent cannot rewrite your profile in that reply.
Everything is checked: each memory must be traceable to something you actually said in a chat, or it is dropped.
Limits: an agent can write a limited number of memories per reply and per day.

You stay in control

One switch for the whole account (Settings → AI agents → Agent memory), and one switch per agent. Off means no tools and nothing added to chats.
Don’t remember this chat — in a chat’s menu; what was learned from it is forgotten too.
Deleting a chat forgets what was learned from it, on all your computers and in the cloud backup. Something another chat also said is kept.
Kept tidy: once a day, memories that say the same thing are merged (“likes coffee” and “doesn’t like coffee” never are), and the memory is packed so it takes much less space and opens faster.
Undo on any “Saved to memory” line, and Download my memory as a file.
Clear memory erases it on every device (you type CLEAR MEMORY to confirm).
Memories nobody has touched for a year are archived and stop being used.
Summaries are written on your computer by default; you can let each chat’s own agent write them instead.

What it can’t do yet

Antigravity gets the memory as text only, without tools.
Search by meaning in the browser is not offered on phones; there it searches by words.