Sleep-time compute: run a second 'sleep agent' to reorganize memory during idle turns

Letta·medium signal

Letta's dual-agent design keeps one agent on live interactions while a second 'sleep' agent activates during downtime to consolidate archival items, rewrite messy human/context blocks, and summarize recent conversation into stable notes. Because memory management happens asynchronously off the critical path, response latency drops and memory quality rises over time instead of degrading. It's a concrete fix for the 'memory gets messy the longer the agent runs' failure mode.