Actor Isolation Must Span Four Agent Subsystems
A technical analysis argues that Actor isolation in agent systems must cover memory, sessions, observation data and scheduled tasks. It says memory may expose preferences and corrections, session records may expose conversation histories and checkpoints, and observation records may contain file contents, command output and tool-call data. Cron jobs require particular ownership checks because an unverified trigger could run under another Actor's identity, write to that Actor's memory and consume its quota. The analysis recommends enforcing authorization at shared write points, treating records without an owner as hidden rather than public, and using compound row identities such as session_id and record ID for session-related records. Its product comparisons and implementation claims are not independently verified in the supplied material.
The analysis says protecting memory alone leaves other sensitive surfaces exposed: sessions can reveal conversation content, observation data can expose files and command output, and scheduled jobs can affect execution identity as well as records.
2026-10-08·记忆系统·Actor 边界
Actor 边界不只是记忆的事。Session 归属、观测数据、定时任务——同一个 actor 字段在四个子系统各有一面。漏守任何一面,泄漏的不一定是"数据",可能是"执行身份"。
一、一个 actor 字段,四个子系统
大多数 Agent 框架在讨论"用户隔离"时只想到记忆。但一次 Agent Run 产生的数据远不止记忆:
| 子系统 | 数据 | 泄漏什么 |
|---|---|---|
| Memory | 偏好、纠正、环境信息 | 个人隐私 |
| Session | 对话历史、消息、检查点 | 对话内容 |
| Observe | Run 追踪、工具调用输入/输出 | 文件内容、命令输出 |
| Cron | 定时任务、执行记录 | 执行身份(以别人身份跑代码) |
#Mem0#LangGraph#Letta#OpenAI Agents#AWS AgentCore#MUXI#Agentium