Accepting defeat on the memory layer

I had pain working with AI. I thought venting to it was a good thing. It made me feel better.

Then I thought, maybe it’s better to build a memory layer. Something that remembers my personal context.

So I started looking around in this lane. I spent months on it. Late nights. I’m not sure how long it actually took. Just to realize a few things.

The first time, I realized the lane is crowded.

The second time, I realized people don’t even need such a thing. Maybe they do. But the nature of AI is stateless. Memory only ever works as instructions. It’ll never be a continuous memory no matter how much you want to fake it.

There are a lot of holes on the development side too. People want an SDK, not a fixed flow. Things like that.

Then today, in a final effort to think through the whole story of memory, I realized something else. Current papers aren’t really solving the root problem, which is that AI agents are still kinda dumb and don’t follow instructions even when the instructions are sitting right there in the context. And I can’t solve that fundamentally. Engineering alone won’t be enough. Probably people at frontier labs are working on it, and I have no close contact with any of them.

Just like that. And I have to admit it. Maybe I wasted a lot of my time doing something completely wrong.

I have a lot of curiosity about the world and I really want to know more about things. But maybe I need to point that curiosity at the correct thing now.

The next fundamental thing is increasing my income by focusing on what I already know. PHP. Web dev. Connecting with people. Trying to sell software, or something like that. I’m not sure yet. But from now on, I’ll do something completely different from investing time into memory for so long.

Peace out.