I got frustrated my AI kept forgetting me so I spent 6 months building a fix

I have a terrible memory myself so maybe that is why it bothered me so much. But every time I restarted an AI agent session it forgot everything. My name, my preferences, what we were working on, all of it gone. Every single time.

I am not talking about within a conversation. I mean across sessions. Restart your agent and you are a stranger again. For a simple chatbot that is annoying. For an AI agent doing real work across multiple sessions that is completely broken.

So I did what any reasonable person does when something annoys them enough. I spent six months building a fix.

The result is Synrix. A local-first memory engine built specifically for AI agents. No embeddings, no cloud, no GPU, no server. Everything runs in-process on your machine and survives restarts, crashes, and power cuts via WAL recovery.

The demo that made me realise I had actually solved something real: I told GPT-4 my name, that I like pugs and Ferraris, and a few facts about my project. Then I restarted the session completely. The GPT-4 instance without Synrix forgot everything instantly. The instance with Synrix recalled every single detail. My name, the pugs, the Ferraris, all of it. Instantly. From a cold start.

That moment is when I stopped thinking of this as a side project.
The architecture is different from vector databases. Instead of embeddings and similarity search, Synrix uses a Binary Lattice structure with prefix-semantic addressing. Lookups are O(k) where k is your result count, not the size of your dataset. Query performance at 50 million nodes is the same as at 1000 nodes. Direct node lookup runs at 19 microseconds. Full agent context restores from cold start in under 1ms.

The honest positioning matters here because I got called out on it early and it made the product better. Synrix is not a replacement for Chroma or Qdrant. For fuzzy similarity search over unstructured documents those tools are exactly right. Synrix is for structured agent memory where you control the naming. User preferences, session context, learned facts, task state. The kind of memory where you know what you are looking for and you need it back in microseconds not milliseconds.

The robotics use case is what excites me most about where this goes. A robot learns its environment during operation. Which door sticks, which patient has a latex allergy, which corridor is slippery after cleaning. Power cuts out. Robot reboots cold. Every memory restores instantly via WAL recovery. No internet required. No cloud dependency. Works in a Faraday cage, on a factory floor, anywhere.

Two weeks after the first Reddit post the project has 80 stars, 600 clones, and 7 forks. Someone running 11 agents on a single Mac mini reached out saying they hit exactly the WAL contention problem Synrix solves. That kind of validation from a real engineer with a real production problem means more than any benchmark number.

If you are building AI agents and hitting the memory problem I would genuinely love to know how you are currently solving it. And if you want to try Synrix the repo is below,

github.com/RYJOX-Technologies/Synrix-Memory-Engine