Ground · Chapter 02
The fact layer
One store of approved facts. Change a fact once and every agent sees it.
The single store of a company's approved facts. Every agent and person reads from it, so a change made once reaches everyone. The ontology keeps it current, not the operator. Glossary ↗
A conversation can contain an old price, a suggested policy and a decision that was later reversed. Remembering the conversation does not tell an agent which statement the business currently stands behind. A fact layer gives approved statements a source, a version and an owner.
The distinction is between a record of what was said and a record of what was decided. Both can be useful, but they answer different questions. When an agent quotes a price or explains a policy, it should be possible to identify the approved fact it used.
One change, shared by every agent
Start with the statements agents rely on most: prices, plans, policies and the description of the customer. Collect them in one place and identify who can approve each one. Give every statement enough provenance to explain why it is treated as current.
Make a small map of the customer, offer, order and policy records before choosing new storage. Reuse the business’s existing source of truth, and give its most important claims an owner, source, version and review date. A notes collection becomes a usable company reference when an assistant can distinguish an approved price from a proposal that happens to mention one.
Agents and people then read through the same access point, whether that is an API or MCP. A change made once should reach every consumer without a separate round of manual corrections. Where an approved fact exists, answers assembled from conversational memory should give way to it.
An agent can still discover a possible correction. That discovery enters review as a proposed fact; it does not become approved just because an agent produced it. Separating proposal from approval keeps the mechanism for learning from becoming a mechanism for silently rewriting the business.
Changing the model as well as the facts
A change to the business’s model can affect more than one statement. The graph-recompile proposal is to rebuild the shared graph and move its consumers to the same version together. This is a stronger claim than saying that every fact needs a version, and it deserves its own test against incremental updates.
Test the reference with a current price, an unknown claim and a contradictory record. The assistant should cite the approved version, identify the conflict and leave the unknown unanswered. Then change the price and follow it to every agent that uses it. If consumers keep quoting different editions, the fact layer has not yet delivered its central promise.
Tools and setup
Existing implementation: Notion documents both assistant access through MCP and AI Connectors. The former connects assistants to Notion; the latter lets Notion search external sources, with capabilities varying by service and plan. Neither automatically establishes which claims your business approves.
- Reuse a source of truth. If using Notion, propose a small facts register with claim, source, owner, status, version and review date. Inspect existing schemas; approve a new schema before creating it in a shared workspace.
- Connect MCP in the setup path in Why a living textbook. Test a known page read. Add an external AI Connector only if the workflow needs it; follow that service's admin/plan requirements and source permissions.
- Write AGENTS.md instructions pointing to the authoritative records, with uncertainty handling and action boundaries. Change a sample approved fact and test retrieval of the new version; invalidate stale working copies.
Setup paths checked against linked documentation on 15 November 2026. Verify current account support during setup; reader integrations have not been tested here.
Work through this with your agent
Paste the request into your assistant. It will interview you and help produce the working artifact.
Interview me to set up a company brain. Ask at most three questions per turn and wait. Start with customers, offers, prices, policies, key entities and who can approve claims. Read authorized existing sources before proposing new storage. Separate raw records, working hypotheses, approved facts, voice preferences and procedures. Draft COMPANY-MAP.md, SOURCE-INDEX.md and a facts register with claim, source, owner, approval status, version, effective date and review date. Mark uncertain or contradictory claims as candidates. Draft an AGENTS.md compatible with existing instructions: source precedence, fact retrieval, uncertainty, approval limits and preservation of history. Agree the destination and review drafts before shared writes. Test answering a price or policy question with the exact approved source/version, then show how a stale or conflicting claim is handled. Do not promote transcript claims into facts automatically. Finish with links and demonstrated versus untested behavior.
Use these documented setup paths as references; check current account support before acting:
https://www.notion.com/help/notion-mcp
https://www.notion.com/help/notion-ai-connectors