I thought LLMs would remember the context, but it turns out they are even more forgetful than a goldfish.

When developing Copilot, I was chatting with the AI through multiple rounds, and finally I couldn't hold back:

It said it had never seen the API documentation I just provided.

It overturned the architecture I just finalized in the next round.

The same bug was fixed three times because it simply didn't remember how I fixed it last time.

That afternoon, I wasted 3 hours repeating "Didn't I just say that?"

Later, I realized: AI is not human; it's a stateless function. If you don't feed it memory, it really is a goldfish.

My approach is simple but effective:

Create a "Project Snapshot" at each stage—current progress, technical choices, pitfalls encountered, and keep it all recorded. The first thing I do in each conversation is feed this to it.

Since then, I haven't freaked out over "Why did it forget again?"

Conclusion: Don't expect AI to be smart. Your job isn't "chatting," it's "feeding memory."