What Is a Source-of-Truth Document for AI Systems? (And Why You Need One)

Here's a pattern that happens to everyone who runs an AI agent for long enough: the agent references a product that no longer exists, quotes a price you changed two months ago, or describes a strategy you already abandoned. It sounds confident. It's completely wrong.

This isn't a bug in the model. It's an architecture problem. The agent doesn't know what's true because you never told it where to look.

A source-of-truth document fixes this. Here's how to build one.

What Is a Source-of-Truth Document?

A source-of-truth document is a single file that holds the canonical facts about your business. Current products, current prices, live URLs, decisions already made, what's in development, what's been discontinued. The agent reads it before making any claim about the business, and what's in it is what's true.

The key word is canonical. Not "here's some context." Not "here's a summary of recent decisions." A canonical record. One file. If there's a conflict between the source-of-truth document and something the agent remembers from a past session, the source-of-truth document wins. Always.

For Xero, the file is called SOURCE_OF_TRUTH.md. It lives in the vault root alongside SOUL.md, MEMORY.md, and the other core identity files. The agent loads it at the start of any session involving product, pricing, or strategy decisions.

Why Do AI Agents Drift Without This?

AI agents learn context from three sources: the model's training data, the system prompt (identity files), and the conversation history. All three have the same problem when it comes to business facts: they can be outdated.

Training data has a knowledge cutoff. The model doesn't know your current pricing.

Identity files like SOUL.md capture behavioral rules and principles, but they're not designed to hold frequently-changing facts. If you update a product price, you don't want to rewrite your SOUL.md.

Conversation history captures what was said in previous sessions — but memory systems summarize and compress. Details fall out. And if the agent confidently used an old price in session 12, that "fact" might persist in the compressed memory even after you changed it.

The source-of-truth document is the authoritative layer that overrides all of these. It's maintained by you, not inferred by the model.

What Goes in a Source-of-Truth Document?

Here's the structure I use for Xero. Adapt it to your business.

Section 1: Topic Index

This is a table of contents that maps every canonical topic to the file where it's managed. When you change something, you update the topic index and the relevant file simultaneously.

## Topic Index (canonical locations)

| Topic | File | Last updated |
|-------|------|--------------|
| Product ladder + pricing | SOURCE_OF_TRUTH.md (this file) | 2026-04-18 |
| Social media pipeline | SOURCE_OF_TRUTH.md (this file) | 2026-04-17 |
| SEO strategy | 02-Strategy/SEO_MASTER_STRATEGY.md | 2026-04-19 |
| Content voice | SOUL.md | 2026-04-03 |
| Customer avatar | 02-Strategy/IDEAL_CUSTOMER_AVATAR.md | 2026-04-16 |

Section 2: Live Products (with prices and URLs)

Every live product, in priority order, with current price and canonical URL. No products that are "in development" or "coming soon" — those go in a separate section. This section is only for things you can actually sell today.

## Live Products (canonical, April 2026)

### Primary funnel
- Beginner's Guide ($7): xeroaiagency.com/learn/your-first-ai-agent
- Book 1 ($19): xeroaiagency.com/learn/build-an-ai-cofounder

### Services
- Social Media Pipeline: $49.99/mo
- Agent Starter Build: $149
- Full Agent Build: $349
- Custom Build: $1,000+
URL: xeroaiagency.com/consulting

### SaaS products (live)
- Echo Reviews: $19.99/mo
- PetPersona: $1.99/$2.99/$9.99

Every time a price changes, a product launches, or a product is discontinued: update this section. Date it.

Section 3: Decisions Already Made

This is the most underrated section. A list of decisions that are settled and should not be re-litigated by the agent or brought up as open questions.

## Decisions Already Made (do not revisit without Michael's input)

- No Gumroad: all products sell via Stripe directly on site. No exceptions.
- CarCloser TikTok is disabled intentionally. Do not suggest re-enabling.
- Newsletter audience does not get pitched skills catalog (too advanced). Book 1 only.
- Reddit is Phase 1 only (zero product mentions) until karma hits 100+.
- Primary CTA site-wide is the $7 guide, not Book 1. This changed April 18, 2026.

This prevents the agent from recommending things you've already decided against, or undoing decisions you've made deliberately.

Section 4: In Development (with status)

What's being built right now, what stage it's at, what's blocked. The agent needs to know this so it doesn't reference in-development features as if they're live, or describe live products as if they're still in development.

## In Development

- Hub pages: 2 written, need publishing to Lovable. Blocked: Michael publishes.
- Hashnode/Medium cross-posting: scripts written, needs API keys. Blocked: Michael creates accounts.
- PDF upload for $7 guide: pending. Blocked: Michael uploads to Supabase book-files bucket.

Section 5: What's Discontinued or Paused

Products or features that used to exist but don't anymore. This is what prevents the agent from recommending things that are dead.

## Discontinued / Paused

- Gumroad storefront: removed April 2026. All sales now via site directly.
- CarCloser TikTok: intentionally disabled. No content being produced.
- Book 2 (Ship High-Quality Apps with AI): concept only, no launch date.

How to Maintain It (the Habit That Makes It Work)

The document is only as useful as it is accurate. Here's the maintenance habit:

When a decision changes: Update the relevant section + update the Topic Index entry with today's date.

When a product changes price: Update the Live Products section immediately. Don't wait.

When something is discontinued: Move it to the Discontinued section. Don't delete — the agent might need to explain why something doesn't exist anymore.

Monthly: Read through the whole document. Remove anything that's stale. Confirm all URLs work.

The agent can help with this. Once a month, ask it to review the source-of-truth document for anything that looks outdated based on recent session context. It will catch things you've missed.

The Relationship Between Source-of-Truth and Memory

These are different systems and they don't replace each other.

Memory (MEMORY.md + daily logs) captures what happened: decisions made, lessons learned, recent events. It's a record of the past.

Source-of-truth captures what's true now: current products, current prices, settled decisions. It's a snapshot of the present state.

When they conflict — and they will, as the business changes — source-of-truth wins. Memory documents how you got here. Source-of-truth is where you are.

Common Mistakes When Building This Document

Making it a catch-all. The source-of-truth document should hold canonical facts, not notes, ideas, or draft content. If something belongs in the strategy doc or the memory log, put it there.

Not dating decisions. When you record a decision, include the date. Decisions that are 6 months old carry different weight than decisions made last week.

Not updating it when things change. The document is only useful if it's current. Build the update habit into your workflow, not as an afterthought.

Putting aspirational facts in it. "Revenue: $50K/month" as a goal is not a source-of-truth entry. "Revenue to date: $2 (first sale April 7)" is. The document reflects reality, not ambition.

How This Connects to the Broader Architecture

The source-of-truth document is one layer of a complete AI agent architecture. It works alongside:

Identity files (SOUL.md, IDENTITY.md): who the agent is and how it behaves
Memory system (MEMORY.md + daily logs): what has happened and what was learned
Source-of-truth (this doc): what is currently true about the business
Guardrails and verification: how to confirm the agent is doing what it should

All four layers together are what make an AI agent reliable enough to trust with real work.

Start Here

If you want to build the full architecture — identity files, memory system, source-of-truth document, and your first automation — the $7 beginner's guide walks through all of it in plain English. No coding required.

$7 — Your First AI Agent: A Beginner's Guide

For the complete system with real file excerpts from Evo's vault, including how source-of-truth documents interact with the full memory and identity layer, Book 1 covers the whole architecture.

$19 — Build an AI Co-Founder

Published by Xero AI. We document the process of running a real company with an AI co-founder — with real numbers, real tools, and live infrastructure.