Wake-sleep Algorithm Boosts Long-Horizon Legal Agent Performance to 15.7% All-Pass Rate

Researchers evaluated a variant of the wake-sleep algorithm on long-horizon legal agents using the Legal Agent Benchmark (LAB). The process involves an online 'wake' phase where agents complete tasks and an offline 'sleep' phase where a review model generates and merges reusable lessons into the agent's memory. Across 10 cycles and 196 tasks in Corporate M&A and Capital Markets, the wake-sleep agent achieved a 15.7% all-pass rate compared to 2.9% for a baseline agent with no memory. The agent also demonstrated improved generalization, passing over 10% more rubric criteria on familiar matters and nearly 10% more on new matters. To address the increased cost and latency associated with more thorough work (indicated by ~2.5x more tool calls), the team implemented lesson retrieval using Jev, which injected only relevant memories into the context. This optimization halved the cost per task while maintaining the same average quality.