Itara

In defense of my laptop's RAM: at least it isn't reading the same billing module four times a day. Sam Altman recently called it "quite terrible" that so much of the world's RAM sits in personal computers that are idle 95% of the time. But somewhere in a datacenter, a GPU is doing exactly that: reading the same file, for the same team, once for every engineer who asks about it. Engineers on a team work in the same codebase, and each of them sends the model the same files and docs. The model processes that context again for every person, on every question, and the company pays for it every time. Itara is a self-hosted gateway a whole team shares. On the team's open model, Llama 3.1 8B on vLLM, it shares the KV cache within the team, so a teammate's question reuses work the model has already done. Other teams can't read that cache or even time it. Llama also reads the repo once and writes a compact memory, stored in Postgres and sealed per team with row-level security, so questions carry that memory instead of raw code. When Llama isn't enough, Muse Spark answers, with the same memory. It all speaks the OpenAI API, so existing tools only need a new base URL. On all 300 SWE-bench Lite issues, Muse with Itara's memory found the right files 86.3% of the time, matching the 86.0% it gets with the full retrieved code, at 57% lower estimated cost. In fast mode, the median answer comes back in 1.9 seconds instead of 12.