I've been away from open source for a while. But the recent explosion of LLMs reignited my curiosity — and one question I couldn't shake:

Why are these models so powerful yet so forgetful?

The answer is in their design. LLMs are trained to be both a reasoner and a knowledge base — fused, frozen at training time. Whatever they knew that day is all they'll ever know. Brilliant reasoning. Zero growth.

When something new happens — a conversation, a decision, a failure — there's nowhere for it to go. The workarounds are painful: retrain the model, stuff the context window, use RAG. None of these are memory. They're duct tape.

So I built Stash. Not to fix the LLM — but to fill the gap it leaves behind. A self-hosted layer that captures what your agent experiences, synthesizes it into a knowledge graph, tracks goals, learns from failures, and reasons across sessions. The model stays frozen. Stash grows.

I know there are other experiments in this space — that's a good sign. The problem is real.…