i built supercompress

i kept running into the same problem with ai agents: context.

when an agent runs for five minutes, everything looks amazing. when it runs for five hours, it forgets what it did, repeats work, loses the goal, and burns tokens carrying around history it no longer needs.

i saw this while building loopy, my autonomous software engineer. the longer it ran, the more obvious the bottleneck became. the agent did not need more prompts. it needed better memory. so i built supercompress.

supercompress compresses long context while preserving the pieces an agent actually needs: decisions, constraints, code references, failures, todos, and current state. early versions either forgot too much or saved too little. finding the line between compression and memory loss was the work.

the system now cuts token usage by about 65% while retaining roughly 98.7% of important information. loopy became cheaper, faster, and able to stay on task longer.

the future of agents is not only better models. it is better systems around the models: memory, compression, state, feedback loops, and execution. supercompress is one piece of that.