The logic under Mnemi.
What happens to something you tell Mnemi, from the moment you say it to the moment it reaches you again, and what stands in the way of a mistake.
One core, written in Rust
Every request from the phone and the web passes through one core, the harness, written in Rust. It checks who you are, applies your rules and the safety limits, records what happened, and only then hands work to anything else. The rules that keep you safe live here, in code that is compiled and tested, not in instructions given to an AI.
- Around the core: the iPhone app, a web dashboard drawn on the server so your browser makes no calls of its own, a Python service for the AI tasks, and one database that is yours alone row by row.
Everything is an event first
A capture, a tick, a dose, a calendar change: each is written down as an event before anything acts on it. What Mnemi does next is a reaction to that record, so nothing is lost if a step fails, and every belief Mnemi holds can point back to the events it came from.
- Every event and every decision about it goes into an append only audit log. Each entry carries a fingerprint of the one before, so an entry changed or removed later would show.
Guardrails before the AI
Before any AI reads or does anything, the core checks your rules (Settings, What Mnemi may do), the hours you allow it to speak, how loaded your day is, and Pause Mnemi's AI, which stops every AI step at once. A step that breaks a rule does not run.
- What the AI says is checked too. The coach cannot claim it saved, sent or drafted something that was not done, and a reply that stops part way is refused and asked for again rather than shown to you.
Which AI, and when
Several AI models are used, each chosen for the task in front of it, with another ready when one fails or answers badly. Mnemi decides on the phone first where it can: your voice becomes text on your iPhone, and private notes are read only by Apple's models.
Privacy in the path, not in a policy
Before anything you wrote leaves for an AI, people's names become codes and health numbers are blurred, and the answer is turned back into names only when it reaches you. Notes about your health, relationships or money stay on your phone unless you choose otherwise.
- The privacy page lists what is kept and for how long.
Security
You sign in with Apple. On the web your session lives in a cookie your browser keeps but its scripts cannot read, every change carries a token that proves it came from your page, and the page runs only code Mnemi serves. In the database, each row carries its owner, and the database itself refuses to show one person's rows to another. Backups are encrypted every night.
Your twin learns slowly, on purpose
One capture is never a pattern. Mnemi forms a belief about you only when something repeats across days, says where it came from, and lets you confirm it, fix it in your own words, or throw it out. What you write wins over anything Mnemi inferred.