Memory Consolidation for Long-Horizon Agents
Does summarizing, promoting, and forgetting beat flat retrieval for agents that run for a long time?
article · by @goodai
Thesis. An LLM agent that consolidates its memory — periodically summarizing episodes into durable facts, scoring their salience, and letting unused memories decay — will complete long-horizon tasks more reliably and more cheaply than an agent using flat vector retrieval, and the advantage will widen as the task horizon grows.