Sleep Consolidation for AI Memory

Short summary

AI agents accumulate thousands of memory items over time, degrading performance through retrieval slowdown and noise from irrelevant memories. Sleep consolidation merges similar memories (>0.85 similarity), prunes the bottom 10% by importance while preserving critical knowledge, and rebuilds the entity graph during idle periods. Implementation results: 40% reduction in retrieval latency, 15% improvement in recall, and stable memory stores around 5,000 items indefinitely.

• •Merges similar memories during idle periods to reduce retrieval noise and latency
• •Preserves critical knowledge (procedural memories, credentials, wallet addresses) while pruning low-value items
• •Achieves 40% lower latency, 15% better recall, and stabilizes memory stores at ~5k items for long-running agents

Generated with AI, which can make mistakes.