Researchers have published a framework designed to stop AI agents from getting stuck in repetitive loops or losing track of their goals during long, multi-step tasks — a persistent problem that has limited the practical usefulness of large language model-based agents.
The paper, posted to ArXiv in April 2025, identifies two distinct failure modes that plague current LLM agents operating in complex environments: what the authors call Progress Drift and Feasibility Violation. Progress Drift describes the tendency of agents to lose sight of the overall goal across many steps. Feasibility Violation describes moment-to-moment errors where agents attempt actions that are logically impossible given the current state of the environment.
Two Failures, One Misdiagnosis
The researchers argue that most existing approaches treat these two problems as a single challenge and attempt to solve them with one unified method. This, they contend, is a fundamental mismatch. Progress Drift is a fuzzy, semantic problem — the kind that language models are naturally suited to handle through pattern recognition and contextual reasoning. Feasibility Violation, by contrast, requires strict logical rules and verifiable state checks, the kind of hard constraints that neural networks tend to handle poorly on their own.
Existing methods typically attempt to address both issues simultaneously using a single paradigm — but these two challenges are fundamentally distinct.
To address this, the team proposes what they call a Neuro-Symbolic Dual Memory Framework — a system that keeps the two problems separate and applies a different computational tool to each.
How the Two Memory Systems Work
The framework operates two memory mechanisms in parallel during inference. The first, called Progress Memory, is neural-network based. It learns from successful past trajectories, extracting high-level "semantic blueprints" that guide the agent toward completing the broader goal. Think of it as the agent building up experience about what a successful path through a task generally looks like.
The second, called Feasibility Memory, is symbolic and logic-based. Rather than learning fuzzy patterns, it synthesizes executable Python verification functions from past failed transitions — moments where the agent tried to do something that turned out to be impossible. These functions then act as hard filters, preventing the agent from repeating the same invalid actions in similar future states.
The two systems run simultaneously, with the neural component steering global direction and the symbolic component enforcing local legality. This division of labour reflects a broader principle in AI research sometimes called neuro-symbolic integration — the idea that combining neural networks with formal logical systems can produce capabilities that neither approach achieves alone.
Benchmark Results Across Three Environments
The researchers tested the framework on three established benchmarks: ALFWorld, a text-based environment simulating household tasks; WebShop, which involves navigating a simulated e-commerce website to find and purchase specific items; and TextCraft, a text-based crafting environment inspired by Minecraft.
According to the paper, the framework outperformed existing competitive baselines across all three benchmarks. The authors also report that it reduced the invalid action rate — how often agents attempt something that cannot work — and shortened the average trajectory length, meaning agents reached their goals in fewer steps. These benchmarks are standard in the field, though results are self-reported by the authors and have not yet undergone independent replication.
The reduction in trajectory length is a practically important metric. Longer trajectories mean more LLM calls, which translates directly into higher computational cost and slower task completion. An agent that reaches the same outcome in fewer steps is more viable for real deployment.
Why Current Agents Struggle With Long Tasks
The problem this paper addresses is not obscure. Deploying LLM-based agents on multi-step tasks — booking a flight, completing an online purchase, managing files across a computer — has proven far harder than benchmarks on simple question-answering might suggest. Agents frequently backtrack, repeat themselves, or pursue subtasks that no longer serve the original goal.
Several research groups have attempted to address this through better prompting, retrieval-augmented memory, or reinforcement learning from feedback. The Dual Memory Framework distinguishes itself by diagnosing the problem at a structural level, arguing that the architecture itself needs to reflect the difference between semantic and logical reasoning rather than delegating both to the same mechanism.
The use of Python functions as the vehicle for feasibility checking is a concrete design choice worth noting. Executable code provides a formally verifiable constraint in a way that natural language rules do not — the agent either passes the check or it does not, with no ambiguity.
What This Means
For developers building LLM agents for real-world tasks, this framework offers a structurally grounded approach to a problem that has resisted simpler fixes — and the combination of reduced invalid actions and shorter task completion paths points toward agents that are both more reliable and cheaper to run.