Has anyone found a good way to stop AI from rebuilding the same mistake twice?

I'm noticing something strange while using Cursor and Claude Code on larger projects.

The first time an AI makes a bad architectural decision, I fix it. I correct the code. I move on.

The second time, I catch it before it happens. I prompt more carefully.

The third time... I realize I'm still fixing the same problem three different ways across three different parts of the codebase.

Here's a concrete example:

I'm building FounderMind AI (a tool for founders to get AI-powered decision insights). I asked Claude Code to implement a new feedback loop feature. It worked. Then I asked Cursor to add a related analytics module. It built something similar, but with a slightly different data structure.

Same problem, two different implementations. Not because the AI was lazy — because I didn't tell either tool about the other's decision.

What I'm realizing:

With vibe coding, you're not just managing code. You're managing context across multiple AI sessions. And each session starts fresh unless you deliberately feed it the history.

The issue isn't that AI forgets — it's that I'm not giving it the full picture consistently.

What I've tried:

• Keeping a running "constraints.md" file I paste into each new session

• Asking the AI to summarize its architecture decisions before I switch tools

• Using the same tool for everything (but different tools have different strengths)

What I haven't figured out:

How to make this sustainable on a 3-month+ project. The context debt grows faster than the codebase.

Question for other vibe coders:

Are you dealing with this too? What's your system for maintaining consistency across AI sessions, especially when you switch between tools?

Do you use a persistent project doc, a specific prompting template, or something else entirely?