The agent frameworks arriving this autumn, Meta’s Muse and OpenAI’s dots among them, are built to finish tasks. Some are also built to stay. frenkit is built for the staying. It makes a bet about what long-term proactive partnerships are made of.
The bet is an idea Bruce Blumberg put me onto at the Media Lab twenty-five years ago, by way of Randy Gallistel’s work on interval timing in animals: a character is made of time. Not calendar time (“what’s next?”) but lived time (“how long has it been? what rhythm are we in?”).
The name is a nod to Pliny the Liberator‘s term of endearment for agent partners, ‘fren.’
A friend notices the topic you stopped raising and the one you cannot stop raising, because they hold the interval between mentions, not just the mentions themselves. A good friend knows when to follow up on one of those topics. So frenkit’s memory keeps that interval: for every person and every topic, when each last came up between us, against how often it usually does, and the gap in either direction is what earns a mention, the way a releaser earns a behaviour in an animal.
The first piece, the ledger and the recall from it, has been running since July on the two characters I live with: one in my family’s group chat that travelled through Japan with us, one at work that watched a contract through its rounds. The part that turns a gap into a mention is still in shadow: it scores, it does not speak, and I read what it would have said. Everything else in the framework is plumbing you could get elsewhere: a borrowed body (it runs on OpenClaw), governed tools, a character file. The brain is the part I am eager to explore, and I won’t know whether there is a contribution in it until it is public.
One of the two, Navy, has already appeared under her own name in AI-Assisted Travel, where I promised the framework for another post. This, finally, is that post.
This weekend I’m running an experiment on proactive support with Navy, at a swim meet. More next week.