Memora — an on-device AI journal that remembers what matters

An on-device AI journal you talk to rather than write in. Tell it about your week the way you would tell a friend; it works out what is worth keeping, and it does that thinking on your phone rather than in somebody's data centre.

The trade every AI assistant asks you to make

An assistant that remembers your life is genuinely useful — the name of your colleague's partner, the restaurant you liked in March, the thing your doctor said. But every mainstream product that offers it works the same way: your words are sent to a server, processed there, and stored under an account you do not control. The more useful the memory, the more sensitive the archive.

Memora removes the server from the equation. A small language model runs on the phone itself, reads what you write, extracts the durable facts, and files them in a local database. The trade — cloud convenience for local privacy — is simply not made.

How the on-device AI journal works

There is no form to fill in and no fields to define. You talk; it files.

• A chat, not a form. The main screen is a conversation. You write what happened, what you are thinking about, what you want to remember, and the model responds naturally — no structured entry, no tags to invent, no folders to maintain.
• Automatic memory extraction. Behind the conversation, an extractor pulls out the facts worth keeping and stores them as discrete memories. There is a heuristic extractor as well as the model-driven one, so the app still records something useful on hardware where the model cannot run at full tilt.
• A timeline you can browse. Extracted memories accumulate into a timeline, each openable in its own detail view. This is what turns a chat log into something you can actually retrieve from.
• Model tiers matched to your phone. A capability detector picks the smallest capable provider for your device, and the model is downloaded once during setup. Weaker phones fall back to a lighter mode rather than being locked out.
• Photos, voice and speech. Images can be analysed on the device using local vision models, and the app supports voice input and spoken replies for when typing is not practical.
• Three languages. The interface and prompts are written for English, Traditional Chinese and Japanese.

Local inference is the whole product

Plenty of apps say “privacy-first” and mean their privacy policy. Here it is an architectural fact: the model weights sit on your phone, the inference runs on your phone, and your conversation never becomes a network request. The only downloads are the model itself, fetched once from a public model host.

That has real costs, and it is worth being honest about them. A phone-sized model is not as capable as a frontier model in a data centre; the initial download is large; and older devices run it slowly. What you get in exchange is an archive of your private life that no company can read, subpoena, train on or leak.

Who it is for

People who tried a cloud AI assistant, found it genuinely helpful, and then thought properly about what they had just typed into it. If that hesitation stopped you using one, this is the version without the hesitation.

It also suits anyone who has bounced off traditional journalling. The blank page is the barrier; a conversation is not. And it suits people who simply want a durable place for the small facts that slip away — medication changes, a friend's new job, what was agreed at a family meeting — without those facts becoming a product somewhere.

Privacy and what leaves the phone

Conversations, extracted memories and photos are stored in a local database on the device. Model inference happens locally. The only network traffic in normal use is the one-time model download from a public model repository, plus the ad and purchase SDKs.

There is no account and no sync, which also means there is no cloud copy — back the device up if the archive matters to you. Memora is free with an optional Pro purchase.

Frequently asked questions

Does my conversation get sent to a server? No. The language model runs on the phone and inference happens locally. The only network download is the model file itself during setup.

Does it work offline? Yes, once the model is installed. Chat, extraction and the timeline all work with no connection.

What if my phone is not powerful enough? The app detects device capability and picks the lightest suitable provider, with a lighter fallback mode instead of refusing to run.

Where are my memories stored? In a local database on the device. There is no account and no sync — which also means no cloud copy, so back up the phone if the archive matters.

Which languages does it support? English, Traditional Chinese and Japanese.