The AI Amnesia Advantage: Why Top Engineers Let Copilot "Forget"

If you’ve ever stared at the VS Code GitHub Copilot documentation and felt completely overwhelmed by .github/copilot-instructions.md, AGENTS.md, and custom agents, you aren't alone.

Recently, I went down a rabbit hole trying to figure out how to properly pass instructions to Copilot. What started as a confusing read through the docs turned into a massive mindset shift about how AI coding assistants actually work under the hood.

Here is a breakdown of the conversation that completely changed how I engineer context in my IDE.

Q: The VS Code docs throw .github/copilot-instructions.md and AGENTS.md at you. Do I need both?

AI: It’s incredibly confusing at first glance, but they are essentially doing the exact same job: acting as a "Rulebook" for the AI.

• .github/copilot-instructions.md is Microsoft’s proprietary rulebook. It tells GitHub Copilot exactly how you want your code written.

• AGENTS.md (placed in your root folder) is an open, agent-agnostic standard. If you use tools like Cursor, Windsurf, or Claude Code alongside Copilot, they will all automatically read this universal file.

You don't need both. If you only use Copilot, stick to the .github folder. If you want your rules to work across any AI editor, use AGENTS.md.

Q: Where do I actually build a custom agent that can use tools, then?

AI: This is where the naming collision trips everyone up. AGENTS.md (capitalized, in the root) is just a passive rulebook. To actually build a custom AI worker with tools, you create a file ending in .agent.md (lowercase) and put it inside a .github/agents/ folder.

VS Code uses a "Split Brain" technique for these files. At the top, you use YAML frontmatter to tell VS Code the agent's name and what tools it can use (like reading files or running terminal commands). Below that, you use standard Markdown to give the AI its persona and instructions.

Q: I used to have this obsession that the AI needed to know absolutely everything about my project context. I wanted it to remember every file, every previous chat, and never forget anything—while still expecting it to write perfect code. Why doesn't that work?

AI: This is the core trap most developers fall into, and it is the big secret that top-tier engineers understand differently. You were suffering from the "Omniscient AI Fallacy." You wanted the AI to act like a senior developer who has been on your team for five years. But in reality, when you force the AI to hold onto a massive, running history of everything you've ever discussed, you trigger Context Pollution. The AI's brain gets too crowded. It starts hallucinating, mixing up the styling rules from your frontend with the database logic of your backend, and its performance drops off a cliff.

Top engineers don't want the AI to remember everything. They ruthlessly clear the context and only feed the AI the exact pieces it needs for the current five-minute task.

Q: So how does the AI actually read my instructions if it doesn't just memorize them at the start of the chat?

AI: You just uncovered the reality of LLMs: AI models have total amnesia. They are completely stateless. Every single time you hit "Enter" in Copilot, VS Code builds a massive, invisible "Hidden Package" behind the scenes. It grabs your recent chat history, the code currently open on your screen, and the specific instruction file for that folder, bundles it all together into a giant wall of text, and sends it to the server. The AI reads it, generates an answer, and instantly forgets you exist.

Q: So the way we talk to ChatGPT on the web is completely different from how we code with Copilot?

AI: The engine is the same, but the packaging strategy is completely different.

When you use a web app, the website acts as a secretary, constantly re-sending your entire conversation history from scratch every time you ask a new question so it feels like a long conversation.

Copilot is a shape-shifter. Because it re-packages everything every time you hit Enter, it can swap its rules on the fly. If you have a Next.js file open, it grabs your frontend instructions. If you click over to a Django views.py file, it instantly drops the UI rules and grabs your Python instructions.

Q: This completely changes my mental model. The AI forgetting context isn't a frustration—it's actually a feature.

AI: Exactly! Every AI has a strict "Context Window" (a limit on how much text it can read at once). We shouldn't treat AI like a human coworker. We need to treat it like a highly specialized, hit-and-run worker.

The ultimate workflow:

1. Open the file you want to edit.

2. Let VS Code automatically attach the dense, short instruction file for that specific framework.

3. Tell the AI to build the feature.

4. Accept the code, clear the chat, and move on.

Give it the blueprint, let it do the job, and dismiss it.

---

This frames your realization perfectly for other developers who are stuck in that same "Omniscient AI" trap.