Inputs and triggers
The daemon can trigger noise generation during a heartbeat. Committed turns can also refresh it, subject to the configured cadence and Ebb behavior. Generation receives selected body, event, conversation, journal, and prior-fragment context. It does not receive every stored memory or every raw tool result by default. The fragment buffer is bounded.Model selection and GPU impact
soma.noise.model selects a dedicated model when configured; otherwise the runtime uses its fallback model. A separate small model can reduce the work per generation, but on a local host it may also require another model load. Measure your actual runtime rather than assuming a second model is cheaper to operate.
The checked-in noise settings enable generation with a 90-second cadence and a 240-token maximum output. The cadence is evaluated at trigger points, so it is not a guarantee of one request every exact 90 seconds.
Optional seeds
soma.noise_seeder can sample inputs such as older episodes, beliefs, knowledge, relationship echoes, journal entries, temporal context, or discovery material. It is disabled in the checked-in defaults. External fetching has a separate opt-in setting.
Seeds are inputs to generation, not verified conclusions. Keep their provenance distinct when fragments later influence beliefs or actions.
Observe and control
ngram seeds and ngram trace expose seed and fragment audit information; use their CLI help for filters. Body inspection shows the rendered fragments available to the Entity.
The shared inference pause blocks GEN model requests along with chat and embeddings. For a runtime that remains active, tune soma.noise.enabled, cadence, token budget, and Ebb settings in the harness configuration.
Source: identity/soma.py and identity/noise_seeder.py.