SafeCommit: Certifying When Memory-Grounded Agents May Safely Act

2026-08-04

AI Summary

This work addresses the problem of memory-driven agents prematurely executing hazardous actions when faced with stale, conflicting, or incomplete memories. To mitigate this risk, the authors propose SafeCommit, a framework that introduces a risk-control layer between reasoning and execution. SafeCommit formalizes, for the first time, the notion of safe commitment under memory uncertainty by constructing a set of plausible worlds grounded in memory, observations, and tool outputs. It leverages conformal prediction to generate action certificates, permitting execution only when an action is deemed safe across all retained plausible worlds. The method disentangles calibration error from representation error to provide theoretical safety guarantees and incorporates probe-action selection with a conservative fallback strategy. Experiments in a dependency-free controllable simulator demonstrate strict probabilistic control over unsafe actions at the target risk level α, effectively balancing safety and utility.

Abstract

Long-horizon agents increasingly use persistent memory and tools to take actions with external side effects. A central failure mode is premature commitment: an agent acts before resolving whether its memory grounding is stale, conflicting, incomplete, or corrupted. We formalize this problem as safe commitment under memory uncertainty and introduce SafeCommit, a risk controlled layer between agent reasoning and external execution. The layer constructs a calibrated set of plausible latent worlds from memory, observations, tool outputs, provenance, and policy constraints. It permits a side effectful action only when a conformal action certificate shows that the action is safe in every retained world. Otherwise, it selects a low-side-effect probe that targets the worlds blocking certification, or returns a conservative fallback. Under calibrated world coverage, the probability of an unsafe certified commit is at most the target level α; with imperfect world proposal, the bound separates calibration and representation error. A dependency-free controlled simulator illustrates the safety-utility tradeoff and reproduces all reported results with one command. The goal is to offer a concrete approach for deciding not only what an agent should do, but when the available evidence is sufficient to safely do it.

Problem

• safe commitment
• memory uncertainty
• premature commitment
• side-effectful actions
• agent safety

Innovation

• safe commitment
• memory uncertainty
• conformal action certificate
• latent world calibration
• risk-controlled execution