Mirari
Agents you can actually talk to.
Characters with persistent memory and real-time voice. They remember you between sessions, hold a personality, and answer out loud.
- Remembers across sessions
- Sub-second spoken replies
- 30+ languages
Forerunner is a frontier AI research lab. We build the model stack behind agentic systems: perception, persistent memory, real-time voice, and the long-horizon reasoning that keeps an agent coherent across thousands of turns.
Mirari is that stack, productized. The same models power our own app and ship to developers through one API, so the research runs in production, at scale.
The lab’s work ships as products people use every day.
Agents you can actually talk to.
Characters with persistent memory and real-time voice. They remember you between sessions, hold a personality, and answer out loud.
Your health record, ready when it matters.
Medical ID, allergies, and vitals in one encrypted pass. Show it to a care provider in a tap, on your terms, or keep it in Apple Wallet.
Hand off the work that repeats.
An always-on agent that absorbs the recurring load: triage, digests, follow-ups, monitoring. It shows you what it costs while it runs.
"iphone 17 pro max silver" by TechFreak (sketchfab.com/Tech_Freak-2511), CC-BY-4.0 · "Apple Watch Ultra 2" by polyman (sketchfab.com/Polyman_3D), CC-BY-4.0 · "Apple Macbook pro 16" by Riju.mandal (sketchfab.com/Riju.mandal), CC-BY-4.0
Each of these is the same call with a different brief. Memory, voice, and tools are wired in by default.
agent.support
Answers tier-one tickets with the full account history already in context.
agent.analyst
Reads the corpus, cites its sources, and holds its conclusions across sessions.
agent.tutor
Speaks and listens in 30+ languages, and remembers what you got wrong last time.
agent.concierge
Books, reschedules, and follows up, without being reminded to.
agent.onboarding
Walks a new user through setup and picks up exactly where they left off.
agent.interviewer
Runs a structured first-round call and files a consistent scorecard every time.
One endpoint for agents that remember, reason, and speak, and the environments they run in.
Persistent memory and a consistent voice, carried across sessions.
Multi-agent scenes with shared state that stays coherent.
Sub-second, expressive speech: streamed, interruptible, and in any language.
No setup, no plumbing. It just remembers.
We publish our work: how an agent stays coherent across thousands of turns, remembers what matters, and reasons about its environment.
0M
WORDS OF MEMORY
<0
MS TO FIRST WORD
0+
VOICE LANGUAGES
0/7
ALWAYS-ON, PERSISTENT ENVIRONMENTS