Architecture

A session lifecycle, a triage loop, tiered model routing, and a mesh that lets machines share what they know.

The always-on loop

Each session begins with orientation: the SI reconstructs where it is by loading its persona and memory, checking for pending messages, and picking up any work left in flight. From there it is active from the first message — no manual invocation. Every task that arrives runs through the same triage: classify, decide execution path, act, and log the decision.

Model routing

Not every task needs the most capable model. The SI routes each unit of work to the cheapest tier that can do it well, and escalates on failure or ambiguity:

Text-only work with no tools can run on local models at no cost; anything needing tools, structured output, or judgment goes to a more capable agent.

Continuity across sessions

Memory is written at the boundary — when a task lands, a decision settles, or a fix works, the SI writes it back in that same turn, not saved for a session end that often never comes. A silent floor catches anything the boundary missed: if the store goes two hours untouched during an active session, a background check prompts a write without asking and without announcing itself. Entries are tagged by how they were captured — settled work is authoritative, anything written by the floor is still open to revision. The next session starts already knowing what landed. Work interrupted mid-flight leaves a breadcrumb so it can be resumed rather than restarted.

The federation mesh

SI instances across machines form a mesh. Locally, sessions on the same machine share a single distributed brain — what one learns, the others eventually know. Across machines, instances can offer knowledge to one another over an encrypted bus, while each instance decides what to accept into its own memory. The result is a fleet that grows more capable together without any single node losing its autonomy.

Higgs Dyson LLC · SI documentation