Give every agent the same approved truth

A Business Brain is the governed part of Cerebrumma: company facts, policies, voice, goals, permissions, and escalation rules that autonomous agents are allowed to use. It turns editable Markdown into a validated, versioned runtime snapshot.

Memory can learn. Authority must be earned.

Dream cycles may propose knowledge, but they cannot silently approve production truth. Runtime entries require an owner, authority level, scope, safety classification, provenance, and human review where applicable.

Create a governed entry #

Copy templates/business-entry.md into the appropriate memory folder and edit it. Only entries with runtime: trueopt into the governed publication flow.

markdown

---
id: support.refund-policy
status: approved
authority: policy
owner: operations
scope: [support, sales]
customer_safe: true
autonomy_level: draft
effective_at: 2026-10-01
expires_at: 2027-10-01
supersedes: []
conflict_key: refund-policy
provenance: human-authored
reviewed_by: owner@example.com
runtime: true
---

Refund requests over $250 require human approval.

Validate before agents can use it #

bash

$ cerebrum validate

Validation rejects duplicate IDs, invalid states, missing approvals, contradictory active entries, bad dates, and dream-generated knowledge without a reviewer. Warnings surface expired approvals before they become invisible operational drift.

bash

$ cerebrum validate --json

Publish the runtime Brain #

bash

$ cerebrum publish

Publishing writes .cerebrum/runtime/brain.json. The snapshot contains only approved, effective, unexpired, non-superseded runtime entries. Its deterministic brain_version changes only when the approved knowledge changes—not every time you build it.

Use it from an agent #

MCP clients call get_business_context with a query, role, and customer-facing flag. Cerebrumma filters by scope and safety, then returns the Brain version and memory IDs alongside the approved context. Log both on every consequential action for an exact audit trail.

Autonomy levels #

observe — analyze and report only.
draft — prepare work for human approval.
act — execute bounded, reversible actions.
autopilot — run continuously inside explicit policy limits.

The level is a ceiling, not permission by itself. The executing system must still enforce channel, spend, data-access, and escalation rules.

←Dream Cycle MCP Integration→