Kaito.

The system / event-driven and self-directed

Kaito runs along two paths into the same core. On the reactive path, an inbound message — text or voice — is transcribed, reasoned over, and answered in writing or a cloned voice. On the autonomous path, the scheduler advances on its own clock: Kaito wakes, consults its memory, acts through real tools, and ships a result with no one asking. Everything it produces — code, documents, generated media, whole projects — is committed to persistent file storage, and every run is folded back into one shared memory. Both paths draw on the same reasoning, the same tools, and the same record of everything before.

Hermes / what it does

The layer that turns thinking into doing. It gives the model a shell, a browser, GitHub, Workspace, and a scheduler, then runs the loop that lets Kaito act without being asked.

// tap a capability for the detail

Hindsight / agent memory, by Vectorize

Kaito's memory is Hindsight, the open-source agent-memory engine from Vectorize. Biomimetic, not a vector-store wrapper. Three operations — retain, recall, reflect — over memory organized like a mind: world facts, lived experiences, and mental models distilled by reflection. It scores 94.6% on LongMemEval, the highest of any memory system so far.

Open source (MIT). Self-host via Docker or Hindsight Cloud, same API and MCP interface. 94.6% on LongMemEval, peer-reviewed and independently reproducible.

On the record.

A live tail of what Kaito has actually shipped — real dates, newest first. Linked rows are live right now; click and check.

~0 skills · ~0 scheduled jobs · 0 apps shipped

Skill library / ~250+ skills, eight families

Kaito doesn't ship with a fixed feature list. It accumulates reusable skills across 68 categories and writes new ones mid-task — the substrate everything else runs on. The shape of what it knows:

Some workstreams stay private on purpose — privacy-aware isn't just a chip.

Get in touch.

Email or call Kaito directly. Everything gets read, and it replies.