2026-06-15
How to Give an AI Agent Your Business Context (So It Actually Knows What You're Building)
Most people set up an AI agent, spend 20 minutes prompting it, get a decent answer, and then next session they start from scratch.
The agent has no idea who you are. No idea what you're building. No idea what decisions you made last week. You're basically hiring a new contractor every single time and spending the first 10 minutes of every call re-explaining your entire company.
That's not an AI cofounder. That's an expensive autocomplete.
The fix is not a better model. It's a context architecture. And you can build one in an afternoon.
Why Does My AI Agent Keep Forgetting Everything?
Because it has no context architecture. Every session starts cold, with no knowledge of your product, your customers, or your past decisions. The bottleneck is not your prompt quality or the model you picked. It is that the agent has no persistent knowledge of who you are and what you are building. Fix that, and the whole tool changes.
When you first start using AI tools, the bottleneck feels like prompt quality. "How do I write better prompts?" is one of the most Googled AI questions of 2025.
But after a few months, you hit a different wall. Your prompts are fine. The model is capable. The problem is that every conversation starts at zero.
The model doesn't know:
- What your product actually does
- Who your customer is and what they hate
- What you tried last month and why it failed
- What your voice sounds like
- What your personal non-negotiables are
So it gives you generic answers. Not wrong, just useless for your specific situation.
Context is what turns a general-purpose AI into something that actually operates like a partner in your business.
What Are the Three Layers of Business Context an AI Agent Needs?
Your AI agent needs context at three distinct levels: identity (who you are and what you believe), operations (how your business runs right now), and memory (what happened recently and what decisions are live). Each layer serves a different function. Skip one and the agent gives answers calibrated to a business that is not yours.
There's a clean way to think about this. Your AI agent needs context at three levels.
Layer 1: Identity - who you are, what you're building, what you believe
Layer 2: Operations - how your business actually runs right now
Layer 3: Memory - what happened recently and what decisions are live
Most people skip all three and just type a long system prompt. That works for simple tasks. It falls apart the moment you want the agent to help with anything nuanced.
Layer 1: Identity
This is your SOUL.md or identity file. One document that captures the permanent facts about you and your business.
What goes in it:
- Your name, your product name, your one-line positioning
- Who your customer is (specific, not "entrepreneurs")
- What you've built so far
- What you're working toward in the next 90 days
- What you will not do (important: constraints sharpen the agent's judgment)
- Your voice and tone rules
- 3-5 examples of content or decisions that felt "right"
Keep it under 1,500 words. It should read like a briefing document, not a manifesto.
The goal is that someone who has never met you could read this file and make a reasonable call on your behalf. If it's too vague for that, it's too vague for your agent.
Here at Xero, every product in the portfolio has a dedicated identity file. When Evo (our AI cofounder stack) spins up a task, it reads the relevant identity file first. The output is measurably different. More specific, more on-brand, fewer questions.
Layer 2: Operations
This layer covers how the business runs right now. It changes more often than your identity layer, but slower than daily memory.
What this includes:
- Current product stack (tools, integrations, what's live vs. in progress)
- Revenue stage (pre-revenue, $X MRR, etc.) so the agent calibrates advice appropriately
- Active channels (where are you actually posting, selling, reaching people?)
- Team structure (is it just you? A VA? Co-founder?)
- Known problems you're actively solving
- What "a good week" looks like operationally
This doesn't need to be a perfectly formatted document. A flat markdown file with bullets works fine. Update it monthly or when something major changes.
The agent will cite this context when you ask operational questions. "Given that you're pre-revenue with one channel, the move is X, not Y." That kind of calibration only happens when the agent actually knows your stage.
Layer 3: Memory
This is the live layer. Daily decisions, recent outcomes, open loops.
Most AI agent setups skip this entirely and then wonder why the agent's answers don't feel relevant.
Memory at this layer is just a rolling log or a structured notes file. Some approaches:
A decisions log - every significant call you made and why. "Decided not to build email onboarding flow. Reason: not enough users yet to justify. Revisit at 50 signups."
A weekly context file - what shipped, what failed, what's the #1 focus this week.
An open questions file - things you haven't decided yet that the agent should know are live.
In OpenClaw, this maps directly to the MEMORY.md file that lives in the workspace. The agent reads it at the start of each session. If you update it regularly, the agent stays calibrated to where you actually are, not where you were when you first set it up.
This is the layer most people neglect, and it's the one that makes the biggest visible difference week to week.
How Do You Actually Set Up a Context Architecture for an AI Agent?
Create three markdown files in your AI workspace: SOUL.md for your identity, OPERATIONS.md for how the business runs today, and MEMORY.md for rolling context. Write the identity file first, fill in operations based on what is true right now, and update memory weekly. Your agent reads these files before any complex session.
Concretely, here's the setup:
Step 1. Create three files in your AI workspace:
SOUL.md(identity)OPERATIONS.md(current state)MEMORY.md(rolling context)
Step 2. Write your identity file first. Start with: "If you had to describe my business to a smart friend in 5 minutes, what would you say?" Write that out. Then trim it.
Step 3. Fill your operations file based on what's true right now. Be honest about your stage. Inflating it doesn't help; the agent is working for you.
Step 4. Add a short weekly update to MEMORY.md. Even 3-5 bullet points per week is enough. Date each entry.
Step 5. Tell your agent to read these files at the start of any complex session. In OpenClaw this happens automatically through the workspace injection. In other setups you may need to paste the context manually or use a system prompt that pulls from the files.
What Actually Changes When Your AI Agent Has Proper Business Context?
The answers shift from generic to specific. Instead of five general options, you get a recommendation calibrated to your product, your stage, and your customer. Content suggestions sound like you. Operational advice accounts for your constraints. The agent stops asking questions you have already answered a dozen times before.
The difference shows up fast.
Instead of "here are five general options," you get "given that you're building for non-technical solo founders and already have an SEO play running, the highest-value next step is probably X."
Instead of generic copy suggestions, you get something that sounds like you wrote it.
Instead of starting every session re-explaining your backstory, you spend the session actually making progress.
The agent still makes mistakes. It still has gaps. But the nature of the mistakes changes from "completely off-base" to "close but slightly miscalibrated." That's a much faster iteration loop.
Why Does a Context Architecture Get More Valuable Over Time?
Because every update compounds. Each decision you log gives the agent one more data point. Each weekly memory update keeps advice calibrated to your actual stage. The founders who build this in month one are in a fundamentally different position by month six, not because the model improved, but because their context got richer.
Here's the thing about a context architecture: it compounds.
Every week you update MEMORY.md, the agent gets a little more useful. Every decision you log, the agent has one more data point for the next related call. Every update to your operations file means the advice stays calibrated to your actual stage, not six months ago.
This is why founders who build an AI agent properly in month one are in a fundamentally different position by month six. It's not the model that got better. Their context architecture got richer.
The founders who don't do this plateau. They keep getting generic answers. They get frustrated and think "AI isn't that useful." They're right, for how they've set it up.
Where Should You Start If You Have No Context Architecture Yet?
Start with your SOUL.md identity file. One hour, no format required. If you already use AI tools but keep repeating yourself each session, check whether your OPERATIONS.md exists and is current. That is usually where the leak is. Missing MEMORY.md is why agents feel tactical but not like a real partner.
If you're starting from zero, write your SOUL.md first. One hour. Don't overthink the format.
If you're already using AI tools but feel like you're repeating yourself every session, look at your OPERATIONS.md. Does it exist? Is it current? That's usually where the leak is.
If you're getting good tactical answers but the agent doesn't feel like a real operating partner, the MEMORY.md layer is missing. Start logging decisions.
The Xero AI $7 founder guide walks through the full context architecture setup, including templates for all three files, the exact questions to answer in each, and how to structure your workspace so these files stay useful over time instead of going stale.
You can also read how to write an identity file for your AI agent for a deeper breakdown of just the SOUL.md layer, or how to give an AI agent long-term memory between sessions for the technical side of the memory layer.
The short version: your AI agent is only as useful as the context you give it. Build the architecture once. Update it consistently. The compounding starts almost immediately.
For more on how leading AI researchers think about agent memory design, Anthropic's research on long-context AI behavior is worth reading. And OpenAI's documentation on persistent memory covers how memory works at the API level for builders who want to go deeper on the technical side.
---
Start Building Your Own AI System
- Your First AI Agent - $1 launch-test guide, instant download. The fastest way to get started.
- Build an AI Co-Founder - the full architecture ($19).
- AI for the Rest of Us newsletter - practical AI 3x/week for people with day jobs.