ZeroSpike

The problem: AI agents resend their entire history to the model at every step: every file read, log and test output. On small machines with small context windows, that history overflows within a few steps, and the model server silently drops the oldest messages first. The oldest message is the task itself, so the agent forgets what it was asked to do, and nobody gets an error. Bigger windows cost more memory and computation, and on hosted APIs every resent token is billed. What we built: ContextGC, a drop-in proxy between any agent and its model (Ollama locally, or Meta's Model API). Before every model call it cleans the conversation history. It removes duplicate messages, moves large tool outputs to disk behind a short pointer, and evicts stale, low-value turns only when the prompt nears the limit. The task and recent work always stay, and nothing is deleted: the original history is kept, so ContextGC can fall back to it if the agent starts looping. Using it takes one URL change and no changes to the agent's code. The outcome: agents finish long tasks in small context windows on small machines, while sending far fewer tokens. On Meta's Model API that halved the cost; on a small local model, more tasks were solved than with a 4× larger window.