AFR: Action-Level Feedback and Cross-Trial Memory for Frozen Language-Model Agents

Large language models are increasingly deployed as agents in interactive environments, where failed actions are frequent. However, two limitations persist. The environment's reply to a failed action is typically a terse no-effect message that carries no repair cue, so the agent repeats state-bound mistakes across trials. Existing remedies entangle two uses of failure information, blocking predicted-to-fail actions before execution and accumulating reflection memory across trials, leaving their individual contributions unclear. To address these issues, we propose AFR, an external interface that decouples the two channels for a frozen agent. Specifically, a state tracker feeds a failure gate that intercepts predicted-to-fail actions and emits a state-bound reason with a repair suggestion, while a memory channel distills each failed trial into lessons injected into later trials. Both attach to the actor at inference time without fine-tuning. On 134 ALFWorld tasks with Qwen3-8B under approximately matched cumulative call budgets, each channel helps on its own: AFR improves success from 61.2% for the bare actor to 88.8%. In a separate evaluation batch, replacing the rule gate with an oracle that consults the environment's admissible-command list increases success from 73.1% to 76.9%, with uncertainty spanning no improvement, while the gate-plus-memory agent reaches 90.3%, above the oracle gate. In a pre-registered transfer test on a second suite, the gate shows no detectable improvement while memory remains beneficial, and a learned-gate checkpoint transfers across actors. AFR reduces avoidable action failures at execution time and reduces repeated failures across trials, without updating the underlying language model; these mechanism-level reductions help explain the success gains under the matched protocol. Code, rollout archives, and analysis scripts are available at https://anonymous.4open.science/r/iclr2027-14BC.