Self-Verifying Scene-Graph Memory for Long-Horizon Embodied Task Execution

Long-horizon embodied tasks expose a persistent weakness of large-language-model (LLM) and vision-language-action (VLA) agents: acting under partial observability and stochastic execution, a reactive policy forgets what it has seen, hallucinates preconditions, and fails silently when an action slips. We present a lightweight agent architecture pairing a structured scene-graph memory—an accumulating, object-centric record of everything observed and every action outcome—with a per-step self-verification loop: before acting the agent predicts the effect of its action, and a verifier compares the prediction against the resulting scene graph, feeding any mismatch back so the agent can self-correct. On a symbolic kitchen/tabletop benchmark with partial observability and a high stochastic slip rate (0.35), and with the policy driven by a free open-weight LLM, structured memory is the dominant factor—raising task success from 0.15 (reactive) to 0.60 and subgoal completion from 0.53 to 0.72—and self-verification adds a further, complementary gain: perfect recovery from stochastic failures (0.901.00), higher subgoal completion (0.720.80), and the lowest hallucinated-precondition rate (0.03). The raw success gain from verification over memory alone is positive but within confidence intervals; its measurable benefit is concentrated in robustness, at a small cost in invalid-action rate, which we report transparently. The system is pure software, needs no GPU, robot, or data, and runs end-to-end on a laptop.