In the beginning there were AI agents — ChatGPT, Claude and the others. Well, actually in the beginning there was the perceptron by Rosenblatt, but let's skip those ancient times.
We all know what an agent is: a stateless function trained on a bunch of data to do things. It's smart, and at this point it can outsmart even an above-average person sometimes. It can do tricks, solve riddles for you, and write code. But we all complain about AI slop, so it's not perfect at all of this.
And here comes the other problem — the long session. Let's say we're working on a project as a developer. The project has a lot of modules and docs, and on top of that there were important transcripts from calls. Oh, and the project is ten years old and some of the technologies are completely custom. But you have an agent now, so things should be easier (and faster) — at least your manager said so, and the head of AI transformation confirmed it.
For some reason, they're not. Agents keep forgetting on long sessions. You're smart, so you use skills, you use a memory.md to get around it, things like a resume — you know the drill. An agent is only as smart as the prompts are smart and the context is there. That's the bottleneck. Big models can hold a lot of context, but context is a raw chunk of data — text, in other words. And you notice it helps, but not completely, so you end up doing all kinds of clever little things to carry it from session to session. We knew this limitation, so we started using vector DBs to store a portion of it. But even then, it's just data. And the difference is called experience.
So here we are. Agents don't have experience. You do.
And what about this Neonmem thing? How is it different from what we already use?
Let me describe it a little. Neonmem is a file — a binary file used as memory for your agent's session. But it's a bit different from what we have now. Inside there's a multi-layer environment: a self-controlled system of layers and nodes that lets the agent have something close to experience, and form its proposals based not just on raw data but using that experience as a grid for its conclusions. The experience grows as you work on the project, with each iteration.
So what about the data, you might ask? It's in there. The file holds a raw-data layer as a mini vector DB, so your agent knows you decided to go in a certain direction — and it has facts to hand you to help you move the way you want.
Next we'll talk about what's inside, and what the cartridge actually is. And maybe a little math as a bonus.
We're at an early stage, but we're moving forward.
Top comments (0)