When Is a Memory Safe Enough to Act On? Action-Conditioned Verification for Long-Lived Agents
Yiqi Wang, Mingkai Zheng and colleagues (University of Southern Queensland, SUSTech, University of Sydney and others) argue that whether a remembered fact must be re-checked depends on the action about to be taken, and build AMV-Guard to decide which memory premises to verify.
Ask this paper
Formulation. Action-conditioned memory verification is posed as a sequential stopping problem that balances execution risk, abstention and verification cost. Verification is worth its cost only when it can change the next action.
Regimes. As the consequence of an action increases, the same memory moves from proceed, to verify, to direct abstention.
AMV-Guard. The gate extracts the premises an action depends on, calibrates beliefs about whether they are wrong, and selects checks by bounded value of information.
Result. On ToolSandbox it lowers verification cost by 33.9% against a static policy with 100% safe-proceed recall and 0% harmful gating, and raises proceed coverage from 38.83% to 44.17%. On AppWorld it lowers cost by 21.3% with a small drop in recall.
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 shipment but should be re-confirmed before a valuable one. We therefore study \emph{action-conditioned memory verification}: given a candidate action, decide which of its memory premises to re-verify and when to proceed or abstain, balancing execution risk, abstention, and verification cost. We formulate this as a sequential stopping problem, in which a verification is worth its cost only if it can change what the agent does next; under suitable verifier reliability and cost conditions, increasing action consequence can move the same memory through proceed, verify, and direct-abstention regimes. Building on this formulation, \textsc{AMV-Guard} is a practical verification gate that extracts the premises a candidate action rests on, calibrates beliefs about whether they are wrong, and selects checks by bounded value of information (VOI). On ToolSandbox, relative to a static policy given the same inputs, it lowers mean verification cost by 33.9\% while keeping 100\% safe-proceed recall and 0\% Gate harmful, and raises Proceed coverage from 38.83\% to 44.17\%; in a gate-core transfer to AppWorld, which carries over the stopping/VOI core but not the full structured pipeline, it lowers verification cost by 21.3\% at a small cost in safe-proceed recall and task success. These results suggest that verification effort should be set by what an agent is about to do, not only by how reliable a memory appears.