The architecture

The Brain holds context. Connected systems hold current operational data. People and AI agents use both through permissions that live outside the model.

1. Chat / email / meetings / documents
2. capture and organise
3. shared memory: decisions, rules, history
4. people + AI agents + skills
5. sales / stock / accounting / other tools
6. verify and keep the result

Keep the boundaries clear

• Capture stores a source reference and its access scope. It does not make every message a fact.
• Retrieval assembles relevant context and reports missing, stale or conflicting evidence.
• A skill describes a method. A capability provides a tested, authorised way to act.
• The source system decides the current value of an order, stock position or invoice.
• A result is complete only after checking its destination and keeping the evidence.

The client and model can change without moving company memory or changing permission rules. A dashboard and an AI conversation can use the same foundation.

First integration test

Use a fictional record. Retrieve permitted context, read the source, prepare a change, stop for approval, apply only after approval and read back the result. Repeat with an unauthorised identity and a failed source. Both must fail explicitly, not fall back to old values.