上下文管理 - 情景记忆

Architecture

从经验中学习模式，其中每次任务运行成为一个情景，包含输入、动作和结果。在处理新任务之前，智能体获取类似情景并将其用作提示。随着时间的推移，智能体感觉像在学习和成长，而不是重复相同的失败计划。

完整 FlowZap 代码

User {
  n1: circle label="1. Submit task"
  n2: rectangle label="12. See result"
  n1.handle(right) -> Agent.n3.handle(left)
  Agent.n9.handle(right) -> n2.handle(left)
}

Agent {
  n3: rectangle label="1. Receive task"
  n4: rectangle label="2. Search similar episodes"
  n5: rectangle label="4. Past lessons"
  n6: rectangle label="5. Execute plan"
  n7: rectangle label="6. Capture outcome"
  n8: rectangle label="7. Send episode for scoring"
  n9: rectangle label="11. Return result"
  n3.handle(right) -> n4.handle(left)
  n5.handle(right) -> n6.handle(left)
  n6.handle(right) -> n7.handle(left)
  n7.handle(right) -> n8.handle(left)
  n8.handle(right) -> Evaluator.n13.handle(left) [label="8. Outcome"]
  EpisodeStore.n12.handle(right) -> n9.handle(left) [label="11.b Outcome"]
}

EpisodeStore {
  n10: rectangle label="3. Episode search"
  n11: rectangle label="3.2 Search similar episodes"
  n12: rectangle label="10. Episode + lesson"
  Agent.n4.handle(right) -> n10.handle(left) [label="2. Search similar episodes"]
  n10.handle(right) -> n11.handle(left) [label="3. Episode search"]
  n11.handle(right) -> Agent.n5.handle(left) [label="4. Past lessons"]
  Evaluator.n14.handle(right) -> n12.handle(left) [label="10. Episode + lesson"]
}

Evaluator {
  n13: rectangle label="9. Score outcome"
  n14: rectangle label="9. Extract lesson"
  n13.handle(right) -> n14.handle(left)
}