Move memory work off the critical path with sleep-time compute / 'dreaming'

Letta·high signal

Sleep-time compute runs a background agent during idle periods that reorganizes context and rewrites the primary agent's memory, turning raw transcripts into learned context before the next query arrives — a Pareto win on latency and cost in Letta's AIME/GSM tests. Anthropic shipped the same idea as 'Dreaming' for Managed Agents: async consolidation modeled on hippocampal replay that merges duplicates and prunes stale entries. If you are not on Managed Agents, replicate it with a scheduled reflection loop that reads past sessions and rewrites memory between runs.