Questions people ask, answered plainly

A memory that does not belong to the model. You tell an AI something once, and every later conversation, in that AI or another one, starts already knowing it. The record is a thing you can open, read word for word, correct, export and delete. Nothing rewrites it behind you.

It is built for the person who named their AI and wants it to still know them tomorrow. Not a database for engineers to wire into an app. There is nothing to install and nothing to keep running.

Odin keeps two ravens. Each morning he sends them out over the world and each evening they return to whisper what they saw. Huginn is Thought. Muninn is Memory. In the Grímnismál he says he fears for Huginn, but fears more for Muninn.

That line carried the whole idea before the idea was found. Thought can be sent out again tomorrow. Memory has to come home. So the product took the raven whose return matters most.

Two reasons, one honest and one lucky. The honest one: Muninn is Old Norse, the root munr (mind, memory, from the verb muna, to remember) plus -inn, the suffixed definite article. The name literally means the Memory. Strip the article and the stem is mun-. That is what is on the sign.

The lucky one: .is is Iceland, and Icelandic is the living descendant of the language the raven was named in. The Grímnismál survives because an Icelandic manuscript, the Codex Regius, kept it. -inn is still the Icelandic definite article; muna still means to remember. The memory raven's name sits on the domain of the language it came from. We would like to say that was planned. It was noticed afterwards, and kept.

Claude, Claude Code and ChatGPT today, and anything else that speaks MCP. You add Muninn as a connector once, pointing at munn.is/mcp, sign in, and from then on the AI has it. The same memory is read by all of them, which is the part their own built-in memory structurally cannot do: Claude's memory never rides into ChatGPT, and ChatGPT's never rides into Claude, because keeping you inside their app is the point of each. Muninn sits outside both.

Theirs is a summary they wrote about you, kept inside their app, revised over time on their schedule. It asks nothing of you and for a lot of people that is enough.

Muninn keeps the record itself, not a summary of it. You decide what goes in. You can read every entry, fix it, scope it to a project, delete it. It comes back exactly as written, and it works in whichever AI you opened. Effortless and theirs, or deliberate and yours. Neither is wrong. They are different deals, and this page is the one where we say so plainly.

No. Nothing is captured behind your back. Muninn holds what you or the AI you are talking to chose to keep, while you were there. If you did not say to remember it, it is not there.

That is deliberate rather than a missing feature. A store that fills itself from your transcripts is a store you have to audit. This one only ever contains what was put in on purpose.

Never. There is no background consolidation, no overnight synthesis, no pass that merges two records into one and hands you the result as fact. What you wrote in June reads back in August word for word.

The cost is real and we take it on purpose: two records can disagree and nothing will notice for you. What that buys is that a wrong record is wrong in the way it was left, where you can see it. A system that reconciles your records is a system holding your pen. Why that matters is a page of its own.

No, and not as a policy but as an inability. Every memory is encrypted before it is stored, under a key that exists at rest only in wrapped form and is unwrapped by the credential you already carry: your password, your API key, your session. There is no master key and no operator key. A dump of the database is ciphertext, wrapped keys and salts, and nothing in it opens anything else in it.

The honest edge: it is not end-to-end encryption, because finding a memory by meaning needs an embedding computed at the moment you save. So plaintext exists for the length of one request, in memory, and is never written down in the clear. The key you hold walks through the whole thing, with the diagram.

Then the recovery code is what gets you back in, which is why it is worth keeping somewhere real. Because there is no operator key, the one thing you can genuinely lose is the last credential that can unwrap your account key. That is the price of nobody else being able to open it either. We would rather say that here than in a support ticket.

Yes, whenever you like. Full export and import, on every tier. You can also have Muninn mirror a copy into your own Google Drive or a GitHub repo as you go, so the memory exists outside Muninn too: readable, diffable, still yours if you walk away. And you can delete any memory, or the whole account and every trace of it, immediately and on your own. A memory you cannot leave with is not yours. It is one you are renting back.

Something a model made that you kept: a document, a diagram, a table, a page. Muninn stores it whole rather than as a link that rots, so it can be made in one model and reopened in another. Because it is stored, it can also be rendered, and a rendered artifact can be published with a single link, like the encryption diagram on this site.

Pages run on an isolated origin that holds no session and no key, so a page written by a model cannot reach anything of yours. That isolation is why the link is safe to send.

Yes, by tag, with a team. You choose which memories are shared; a teammate's AI sees them transparently, and there is one source rather than copies. Under the hood the team gets its own key, wrapped once for each member under that member's own key, so nobody hands anyone their personal key to work together. Remove a member and the team rekeys; the copy they could unwrap is simply dead.

It does not consolidate, and that is absent rather than deferred. It does not guarantee total recall: it surfaces what is relevant by meaning and by how often you lean on it, not every fact you ever gave it on cue. Search is eventually consistent, so a memory you just added can lag a few seconds before it is findable. It does not infer things about you unprompted; if you want a memory that quietly writes a profile of you, Claude and ChatGPT do that and Muninn does not.

And it is early. It is in daily use by one person, on real work, and the honest state of it is written on the product page. A tool that promises perfect recall is selling the comfortable version nobody can check. This is the one that can be.