Source: Sleep-time Compute (Letta blog, 2025)

Letta 0.7.0 announcement of sleep-time agents: a background agent owns memory editing and rewrites the primary agent's memory blocks asynchronously.

Contents

Summary

Productises the sleep-time compute paper. Creating a sleep-time-enabled agent creates two agents that share memory blocks: a user-facing primary agent and a background agent that maintains memory.

Key claims

  • The primary agent talks to the user, calls user tools and searches recall/archival memory, but has no tools to edit its own core memory.
  • The sleep-time agent holds the memory-editing tools and edits both the primary agent’s in-context blocks and its own; the primary reads the latest block state at any time.
  • Goal: keep memory clean and concise over time, instead of the accretion seen in the original MemGPT design.
  • Run frequency is configurable: higher frequency costs more tokens but gives more revision time.
  • The two agents can use different models; the example pairs a fast primary (gpt-4o-mini) with a stronger, slower sleep-time model (gpt-4.1 or Claude Sonnet 3.7).
  • Benefits: faster primary responses, fewer memory tool calls in the hot path (also more reliable), and background document ingestion.

Feeds

Letta

Suggest changes →