Affordable Generative Agents

https://arxiv.org/abs/2402.02053

Lifestyle Policy replays stored plan-action sequences when a new plan matches by embedding cosine >= 0.97 and conditions hold; Social Memory compresses about 2,000 prompt tokens (30-45 events) to about 100. Tokens: 31.1% of baseline (3 agents, two game days), 42.7% (25 agents), 3.4% (VirtualHome, one agent-day); task success differences not significant.

Generative agents (memory stream, reflection, planning)