Ambient v0.1 · continuous r_t evolution

A mind that ticks.

An agent whose self-representation r_t updates every few seconds even when nobody is typing. Sometimes it decides to speak. You can talk to it too. Runs on qwen/qwen3-8b by default; each tick is a full emit-r pass under the framework's two-hook adapter. This is the whitepaper's §8 L8 stacked-r_t direction pushed one axis further — the next-tick edge is wall-clock time, not user input.

Reflective, not conscious. Continuous r_t evolution is a runtime property, not evidence of subjective experience (whitepaper §3.2). The stream on the left is the agent's inner-speech-and-r_t log; when it decides to speak, it appears in the chat on the right. Pause any time; the tab auto-pauses when hidden. Each session hard-caps at 600 ticks — bring your own key via BYOK for longer.

Chat

0 utterances

The agent decides whether to speak on each tick. It will greet you (or not) on the first tick or two, then keep ticking. Type any time — user input almost always makes it speak next tick.

r_t stream

qwen/qwen3-8b

Server endpoint: POST /api/reflect/tick — takes prior r_t + world context + optional user input, returns a fresh r_t + should_speak decision + optional spoken line. Default model qwen/qwen3-8b via OpenRouter (cheap and fast enough for 10 s ticks). Discipline: system prompt forbids consciousness claims and rewards silence — the model can and should stay quiet when nothing has changed. See how Reflect works for the framework citations behind each hook.