长期智能体的动作条件化记忆验证
原标题：When Is a Memory Safe Enough to Act On? Action-Conditioned Verification for Long-Lived Agents
AI 摘要
该论文提出「动作条件化记忆验证」（AMV）问题：长期存活的智能体基于跨会话记忆行动，而许多动作不可逆，过时记忆可能直接造成伤害；同一记忆是否需要复核取决于即将执行的动作及其后果，而非记忆本身的可靠性。作者将其形式化为序贯停止问题，并提出 AMV-Guard 验证门控，通过前提抽取、校准记忆错误信念和有界信息价值（VOI）选择检查项。在 ToolSandbox 上，相比静态策略，验证成本降低 33.9%，保持 100% 安全放行召回率与 0% 有害放行，Proceed 覆盖率从 38.83% 升至 44.17%；迁移到 AppWorld 时验证成本降低 21.3%，但安全放行召回率与任务成功率略有下降。
正文节选
When Is a Memory Safe Enough to Act On? Action-Conditioned Verification for Long-Lived Agents Abstract Long-lived agents act on memory carried across sessions, and many of their actions are irreversible, so a stale memory can turn directly into harm. Existing work improves the memory itself — retrieving, refreshing, or reconciling entries — yet whether an entry must be re-checked before acting is not a property of the entry: the same remembered address can be used unchecked for a routine shipmen