智能体 / NAVIN RESEARCH
AI Agent Memory: What to Store and What to Recompute
Separate working context, durable user facts, episodic history, and external knowledge to reduce cost and privacy risk.
Separate working context, durable user facts, episodic history, and external knowledge to reduce cost and privacy risk.
简要回答: Treat memory as several stores with different lifetimes. Keep task-local state in working context, verified preferences in a durable profile, reusable facts in a governed knowledge store, and execution traces in observability storage. Do not call an unfiltered transcript memory.
实施步骤
- Define a purpose, retention period, and deletion path for each memory class.
- Write only facts supported by user confirmation or trusted system events.
- Retrieve by task relevance and recency, then cap the amount inserted into context.
- Allow users and operators to inspect and correct durable memory.
验证清单
- Test whether stale facts are superseded correctly.
- Measure retrieval precision, not only recall.
- Verify tenant isolation and deletion across backups and indexes.
常见错误
- Saving every message forever.
- Letting model-generated guesses become durable facts.
- Mixing audit logs with prompt-ready memory.
实际决策
先交付能够通过上述检查的最小设计。记录测试中使用的模型、提示词、工具、数据集与运行时版本，让后续变化可以被比较而不是被猜测。依赖供应商的限制、模型名称、价格或预览功能之前，请重新检查链接文档；这些细节变化往往快于本文的工程原则。
