Keep many coding agents moving with local context cache

A developer ran several Claude Code sessions at once on local RTX 3090 hardware and avoided reloading long contexts after idle periods.

Done with Claude Code· Advanced· Oct 2026

What they did
Roark66 paired Claude Code with local models on RTX 3090 cards that each had 24GB of memory. They kept as many as 10 sessions open, though only about 5 were doing work at any moment. Since they controlled the inference stack, they preserved entire contexts in RAM and could write them out to disk, so an idle session did not need to start over from scratch.
How it went
They said keeping roughly 15 complete contexts in memory drove cache usage to 95% or higher. They also estimated that their weekly usage of about 1.3B input tokens would run into thousands of USD if billed at API rates.
Tips
This approach depends on having enough RAM to hold multiple long-running sessions. It also sidesteps cloud cache expiry, which can make an idle long context expensive to resume.

Try it yourself with Claude Code

I need to handle [N] independent coding jobs in [path] at once. Split them into separate git worktrees and branches, keep each one isolated, run the needed tests, and give me a brief status update for each task: [task 1], [task 2], [task 3]. If you reach a step that would delete files, send changes, merge branches, or otherwise commit to something I may not want, pause first and ask for my approval. Do not combine the branches unless I explicitly say so.

Fill in the [brackets]. Agents can make mistakes, so review anything before it is sent, paid or deleted.

Read the original on news.ycombinator.com ↗

Get 3 agent uses like this every morning

With the day's top AI news. Free, and one click to unsubscribe.