What Our Own Memory Layer Changed About How We Build

What changed once we stopped re-explaining context to our own agents every session, and built something to fix it.

As a software engineer, I am consistently looking for better ways to improve how I work. When AI became a useful tool in software development, I wanted to understand how it could improve my workflow. Though I discovered ways AI could help me become more efficient, I also ran into things that caused frustration. The biggest frustration was managing context. I would have to constantly catch my agent up before it could be helpful. Although AI has improved at reading a codebase, there was still so much about a client or a product's strategy AI didn't know unless I provided it. I tried handing it vision docs and specs for projects, but that created more confusion. These documents are usually poorly written, or the client hadn't fully decided what they wanted. The AI was left to determine the scope and direction, and know what decisions had been made and why. My context window was flooded with content that wasn't helpful. I needed a way to provide relevant context without managing all of it manually.

Around the same time, my business partner Chris and I noticed the same pattern with our clients. Whether they were writing content, building their strategy, preparing for meetings, or drafting prototypes to help communicate their needs, the story was the same. Day after day, their AI agent seemed to keep forgetting what they were working on. They would start a new conversation and re-explain where they left off.

The shared frustrations experienced by our clients and other engineers led us to build SpaceStationAI. SpaceStation is a context harness created to solve three problems: give agents curated context instead of a document dump, stop re-explaining our own preferences every session, and control what context is relevant to what we're working on at any given moment. SpaceStation didn't come together overnight, though. Chris and I explored a lot of existing ideas first. We liked the "second brain" concept, shared context for all your agents, but it didn't solve our biggest problem…