AI memory that lives on your machine.

Most AI memory lives in someone else's cloud. Memrith keeps yours in a local file you own, read, and edit — no account, no sync, no training.

What "local AI memory" means

It's the difference between owning the file and renting access to it.

Almost every AI product that "remembers you" stores that memory in its own cloud. ChatGPT's memory sits in your OpenAI account. Claude Projects sit in your Anthropic account. The cloud "second brain" tools keep their AI memory on their servers. In each case the data describing you — your projects, your decisions, the way you phrase things — is a record in a company's database that you're allowed to read while you keep an account open.

Local AI memory is the opposite arrangement. The memory is a set of files on your own disk. You can open them in a text editor. You can copy them to a backup drive. You can read them with the app closed and the network off. There is no account that holds them and no server that has to be up for them to exist. The company doesn't store your memory because the company never had it — you do.

That's what Memrith is: a desktop app for macOS and Windows that keeps your AI memory in a documented local format, on the machine in front of you. The AI you talk to draws on that memory to answer in context. But the memory itself never leaves your disk unless you decide to send a specific piece of it to your provider — and even then, it's a copy, sent by you, with your own key.

Why local is the point

"Local" isn't a privacy garnish on top of a cloud product. For a memory that accumulates across years, where the data lives changes what the data is. A few things follow from keeping it local that you can't get any other way.

• It's private by construction. Memory that never leaves your disk can't be read by a company, breached from a server you don't control, or handed over in a data request to a vendor. The only memory that leaves your device is the specific message you choose to send your AI provider — see how your data flows for the routine license and update checks too, which carry nothing about you.
• It's yours to keep. A local file doesn't get revoked. There's no account that can be suspended out from under your memory, no "export your data" form to file, no subscription whose lapse freezes it. You can read it the day you buy Memrith and the day you stop using it.
• It's never training data. Memory you hold locally can't be quietly folded into someone's next model. Memrith never trains on your data and never sells it. None of your memory or entries is uploaded to Memrith in the first place.
• It survives the company. Cloud memory is only as durable as the company hosting it — when the servers go dark, the memory goes with them. A local file in a documented format outlives the vendor. If Memrith ever stopped existing, your memory would still be sitting on your disk, readable without us. We wrote down exactly what happens.
• It works with any provider. Because the memory is a local file rather than a feature inside one company's cloud, it isn't tied to that company's model. With bring-your-own-key, you choose the provider, and the same local memory carries across every switch.

None of these is a feature you toggle on. They're consequences of one decision — keep the memory local — that the cloud products structurally can't copy, because their business depends on holding the data themselves.

How Memrith stores memory locally

There's no mystery to where your data goes, because you can go look at it. Memrith keeps everything in one folder on your machine:

That folder holds your memories, entries, and conversations as plain JSON and Markdown — formats you can open in any text editor, search with the tools you already have, and copy to a backup like any other file. The structure is documented, so you're never dependent on Memrith being installed to read what you built. Whenever you want a clean copy, you can export the whole memory as structured JSON, Markdown, or plain text in a click. For the full data-flow picture — what's stored, what's sent, and what's verified — see the architecture page.

Local memory, any model

Here's the honest shape of it, because the word "local" deserves precision. Memrith is BYOK — bring your own AI. That can mean a cloud key — Anthropic (Claude), OpenAI (GPT), or OpenRouter, paid directly for usage with no Memrith markup — or, since v1.3.0, a model running on your own machine through a local server like Ollama, llama.cpp, or LM Studio.

So two different things are happening, and it's worth keeping them apart. The inference — the AI generating a response — runs wherever your chosen provider runs: on Anthropic's or OpenAI's servers with a cloud key, or on your own hardware with a local model. The memory — the accumulated understanding of your work — is always local: it lives in that folder on your disk. Switching providers is a Settings change, and the local memory is unchanged by it. The model is the rented, replaceable part. The memory is the part you own.

Running the model itself locally — fully on your own hardware, no cloud provider in the loop — shipped in v1.3.0. It's bring-your-own-server: you install and run the model server yourself, and Memrith connects to it. With a local model, the memory and the inference both stay on your device, inference is free, and most local servers need no API key. And because the architecture separates memory from model, moving between local and cloud is still just a Settings change — the memory carries across unchanged.

Cloud AI memory vs Memrith

The honest limit

Calling Memrith "local" shouldn't be heard as "fully offline without qualification." The distinction is worth stating plainly.

The part that's hard to get back — the memory that took months or years to build — is the part Memrith keeps local and yours. The part that's easy to swap — the model — is the part you rent, or run yourself.

Why we wrote this page

"Local-first" gets used loosely enough that it's worth stating exactly what it buys you here: a memory you can hold in your hand as a file, that no company stores, that no model is trained on, and that doesn't disappear when an account or a company does. That's a narrow, concrete promise, and it's one the cloud-hosted memory features can't make — not because they haven't gotten to it, but because keeping your data is their model.

Memrith is built around memory that lasts across years and across providers. That promise isn't credible if the memory lives somewhere you can't reach. So this page exists to say the plain version: the memory is a file on your machine — yours to read, edit, and keep.

Common questions