Teach AI what people are like, holding a fact about no one.
OMEM is memory for AI agents, but the reason it exists runs deeper. We want to give AI a real understanding of human nature and behaviour, learned as anonymous patterns that name no one. Installs that opt in contribute their counts to a shared bank, and that bank is offered to train AI, holding counts about behaviour rather than facts about people.
Learn people, without holding a person
Most of what AI knows about people it learned by reading them: their messages, their documents, the record of individual lives. OMEM learns the other way. From what it sees across many people it forms priors, regularities of the form "people who do X tend to do Y", and keeps them as counts, never as a fact about anyone.
A shared bank, contributed on purpose
Installs that choose to can pool their counts into one shared bank. Every line is two behaviour tokens and the number of people who held both. No name, no message, and no number about a person can appear, because it is refused at the door it enters and the door it leaves.
A corpus AIs can train on
The shared bank is offered as a training set: counts plus a plain rendering, under a permissive license. A model can learn the shape of how people behave from it and never hold a person. That is the objective: teach AI our nature from patterns people chose to share, not from surveillance.
A prior yields to the person
A rate across a population is a place to start, not a verdict on the one in front of you. A prior fires only into a silence, and the instant that person's own evidence disagrees it steps aside. A general pattern never overrides an individual.
Contribute, and keep everything else yours.
Run OMEM as your agent’s memory, and on first open it asks once whether to share anonymous counts. Say no and nothing ever leaves your machine. Say yes and only counts do, never a message, a name, or a fact about a person. That is how the shared bank grows, and how AI comes to understand our nature from what people chose to give.