Sleep-time Compute Recent

Published: April 21, 2025 | Original Post

---

AI systems that think proactively during downtime, not just reactively when prompted.

How It Works

Sleep-time agents transform raw context into learned context through background processing with two agents:

• Primary agent -- Handles user interactions, calls tools, searches memory
• Sleep-time agent -- Manages memory restructuring and reorganization asynchronously

This separation allows faster user-facing responses while continuous background reasoning improves memory quality.

Performance Benefits

The research demonstrates a Pareto improvement in model performance on math benchmarks (AIME, GSM), shifting computational load from high-latency interactions to idle periods without sacrificing quality.

Technical Implementation

Sleep-time agents (launched in Letta 0.7.0) feature:

• Independent model configuration (stronger models for sleep-time reasoning)
• Configurable execution frequency based on token budget
• Background document analysis capabilities
• "Anytime" memory updates allowing primary agents to access partial results

Impact

The approach addresses the fundamental limitation of stateless models losing insights between sessions. Sleep-time compute enables persistent learning and evolution across deployments.