Hermes learning sidecars
A process-learning observer for each active holon, without giving the observer authority.
Each active non-principal holon can have exactly one logical Hermes sidecar. Hermes observes outcomes, harvests evidence, digests repeated patterns and promotes evidence-backed lessons. It does not consume a Rule-of-Three child slot and cannot dispatch, merge, deploy or widen authority.
Hermes is the team's notebook-and-coach, not the boss. It watches what worked and failed, notices repeated mistakes, and suggests lessons. A lesson only travels to other teams after it is deliberately promoted with evidence.
Hermes closes the process-learning loop without letting an AI memory system rewrite policy. Raw observations stay local. Promoted lessons have evidence references and an explicit direction (local, up, down or both). If Nous Hermes Agent is used as the execution provider, each holon gets its own isolated profile so automatic memory writes cannot contaminate another holon's state.
Sidecar, not child orchestrator
SYNC → RECORD → HARVEST → DIGEST → PROMOTE
Lessons have direction
Raw observations never propagate as truth. The promoted lesson is the controlled boundary.
| Scope | Visible to |
|---|---|
| local | This sidecar only |
| up | Eligible ancestors |
| down | Eligible descendants |
| both | Eligible ancestors and descendants |
One isolated profile per holon
- Use one sterile template profile named primitive-hermes.
- Generate deterministic per-holon profiles from the template, not from the parent profile.
- Do not point two agent processes at the same Hermes profile.
- Treat profile memory as execution state, not authority state.
- Cross-holon learning still travels through promoted Control lessons, not raw private memory.
When to run Hermes
The Control runbook calls for sync at orchestration-cycle start, harvest after worker-ledger collection, and digest before changing a repeated process. This makes learning continuous while keeping policy changes explicit.
Related pages
Canonical sources
These pages are explanatory read models. Where wording conflicts with implementation or a Control decision, the linked implementation/decision is authoritative and this page should be corrected.