All demos

Waiting agents give their memory back

Shared by Daniel Sadilek on November 9, 2026
Area Agents & modelsKind Performance

An agent that waits for the owner's answer (a card's worker, an idea's agent, a project agent) used to keep its process and several hundred megabytes of memory for as long as the owner took to reply. Now the process ends once the agent has waited longer than its prompt cache lives, an hour for Claude Code, and the owner's next message resumes the conversation where it stopped. Measurements show that this costs no extra tokens: after an hour, a process that was kept writes its whole conversation to the cache again anyway, and a resumed one writes the same amount.