A little background

About a year ago, I started coding regularly with AI coding agents and found the experience to be 2 parts exhiliration and 1 part frustration.

I'm pretty steeped in .Net at this point in my careear. So, for fun I tried writing a couple applications in Typescript and Rust. I used a mix of Claude Code, Copilot CLI, Gemini CLI and Codex. I was honestly pretty blown away by how quickly AI helped me assimilate new languages.

It wasn't all a joy kindling experience though.

At first, I didn't understand the context window, how to manage it, or how working with agents is like working with amnesiacs.

If I didn't know how common it is, then I'd be embarrassed to admit that I found myself cursing at my screen on more than a few occasions, but the truth is that I did.

I began to figure out the context window, but remained frustrated that the agents didn't remember decisions 'we' made.

My first attempt at achieving continuity across sessions yielded a system that I think many have stumbled upon (the session dump). My diary of sessions began to grow. And it worked great until eventually all the embedded information was doing more to distract the agents than keep them aligned with my intentions.

I decided that I needed something better, and that is when Jumbo was born.

The project got its name, because I thought I was setting out to build memory for coding agents. There's a trope about elephants never forgetting, and so an elephant named Jumbo seemed like a good mascot.

Since I was building memory for agents, I thought it would be wise to understand how memory works in the human brain, and started doing some reading. I found out that, through pure intuition, I had built a system that closely models the processes involved in working memory. Working memory is the function in the brain that allows us to accomplish goals. It's dependant on long-term memory, and you're ineffectual without it.

[A quick aside for anyone interested in the subject, or maybe building your own memory system]

My revelation came from a book entitled 'Permanent Present Tense' by Suzanne Corkin. She writes about a neuroscience case study that perfectly captures the frustration of working with AI coding agents.

Henry Molaison had portions of his hippocampus removed to treat epilepsy. He retained all his existing skills and knowledge, but after to the operation lost the ability to form new long-term declarative memories. He could act, but couldn't remember facts or events. He was competent — but perpetually starting over.

That's the AI coding agent problem in a nutshell.

What I built

After months of dogfooding my own approach, I released Jumbo CLI — Open Source Memory and Context Orchestration for Coding Agents (Claude Code, Copilot, Gemini, etc.).

The project evolved into more than a bolt on memory system. Its a platform that orchestrates the management of my context window for me.

What makes it unique is the goal primitive.

Without goals, a memory base is basically just a …

为什么值得关注

提供了用户原本不知道的新信息；能改变理解方式，而不只是重复常识；它带来了新的事实、进展或信息，不是在重复旧内容

来源：reddit，领域：projects，保留分：0.69