Feedback To Devs

Author: jryanhaberCreated May 5, 2025Updated Aug 26, 2025

Great start.

I gave this a run with a solid project I am using and ran into the same problems as when I built my own memory system. Token Usage Capacity, which actually made it not usable in any fashion in its current state as a drop in replacement for claude code, which had been my hope.

My sense is this: theres a ton of optimization for token usage that prevents claude code from maxing out your tokens on a single request... and your system would need such optimizations to actually have workflow value, at least in contexts like mine.

Im willing to entertain if I am wrong about this, or if this is not intended for production environments, or not intended for real world use cases, or just a wip...

But if it was me, I would absolutely ask myself deeply what assumptions I made in the context window utilization efficiency for this, and double check each of those assumptions. There are likely examples of how claude code handles this on github somewhere, and that looks like the next and most critical step for this.

My reasoning:

I have a very very mature workflow with claude code agents, and had tried building my own memory via mcp, it literally just maxed its context window on the first call to it and I had to shut it down.

optimizing my workflow has meant getting in there and injecting the right minimal context at the right moments, I use text expander for that.

With your system, same problem. I literally just asked the agent what paths it had access to, and it pumped its context window completely full ( 20k tokens ) and shut it self down. thats the same issue I had.

Not sure how to get value from it in the current condition, but I was super curious about your progress because you are pushing an edge that I think does need to be pushed... we just need to get in there at each step of the inner mechanics of its process and LOCK DOWN all of the context consumption / injection to make sure its clean and minimized, and perhaps use cheaper models for specific aspects of the workflow.

Love your direction with this, hope you continue.

Source: Doriandarko/claude-engineer