← Notebook

Moving the life inside the model

There was a structural ceiling in the early Path B runtime, and on 2026-05-24 we set about removing it.

The ceiling

Until now the model was called. Each moment of the organism’s life was a fresh request: tokenise, run the model, get an answer, discard. Even with caching for speed, the shape was request-and-respond. Two things follow from that shape, and both are fatal to what we are trying to build.

First, identity has to be reconstructed every time from a text transcript, because nothing of the model’s actual internal state survives between calls. Second, and worse, a thing that only answers when asked cannot be the active agent of its own life. It halts the instant it stops generating. “Thinks on its own” is not a tuning problem in that design; it is architecturally impossible.

The native runtime, built in rungs

We built a new runtime as a ladder of independently shippable rungs, and landed the first three:

  • Held state across moments. The model is loaded once, inside our own process, and its internal generation state is a live object we keep across ticks, not rebuilt from text each time. We can write that state to disk and read it back later; a life becomes a file plus a running loop.
  • Generation that does not stop at an action. The model emits a continuous stream. A watcher reads actions out of the stream as they appear, dispatches them to the body and world, and lets generation continue. The organism can narrate, act, change its mind, and narrate again with no intervening call.
  • The world spliced in mid-thought. Sensory updates are injected directly into the running state as inline markup, so the organism perceives a change in its world from inside its own ongoing thought rather than as a fresh prompt. The runtime stops the model for nothing except death.

The architectural shift is the whole point: the runtime is no longer something that calls the model. It is the body and world the model lives inside.

The organism's within-life continuity is now carried by the model's own retained internal state, held in-process and persisted to disk, not reconstructed from a transcript.

Not: holding that state is an engineering fact, not a claim that the state is a mind or that the organism is aware of its own continuity. It says the continuity has a real physical location we control and can verify; what that continuity means is a separate argument we make only as far as the operational framing permits.

What is deliberately left for later

There is a fourth rung we did not build: decoupling the world’s clock from the model’s clock, so that time becomes a genuinely independent dimension: the world able to move while the model is mid-thought, the model able to think while the world holds still. It opens real design questions and we have parked it as the next research direction rather than rushing it.

What this day bought is the precondition for everything that follows: a life that is a held internal state on disk, resumable later, with the model as the active thing and the substrate as the world around it. The very next entry is the test of whether that “held state on disk” is as literal as the architecture assumes, or whether we have been telling ourselves a comfortable story.