记忆缺口：为什么你的智能体会忘记刚学到的东西

AI agents can tackle complex problems, plan, execute, and iterate, yet they often fail to recall actions taken just minutes prior. This issue is not a bug but rather an architectural oversight that hampers team productivity. Most agent frameworks treat memory as an afterthought, relying on context embedded in prompts and hoping the model retains this information. However, this approach typically fails when conversations exceed a few dozen turns. Recently, the ghost team developed an intriguing solution: ephemeral Postgres databases for agents, as opposed to vector stores or embedding databases. These relational databases can be spun up in seconds and vanish after use.