What Makes Agent Memory Useful for Reliable Unanswerable Question Handling?

Chuanyuan Tan, Junjie Yu, Yuxin Wang, Yining Zheng, Xipeng Qiu, Wenliang ChenarXiv 2026 · arXiv

TL;DR

This paper presents a systematic study on how external memory helps LLM-based agents handle unanswerable questions (UAQs). The authors discover that memory gains are highly selective and fragile under dataset shifts, but surprisingly robust across different base models. Ultimately, memory is most effective when it provides high-level procedural and rule-based decision guidance rather than just altering step-by-step tool interaction trajectories.

What Problem Does This Paper Address?

Large Language Models (LLMs) frequently struggle with unanswerable questions (UAQs)—queries that contain false premises, non-existent entities, or suffer from insufficient evidence. While standalone LLMs can be trained or prompted to say "I don't know," the problem becomes significantly more complex in agentic settings.

An LLM agent interacting with an environment via tools cannot just give a flat refusal; it must make sequential decisions. It has to determine whether to keep searching for evidence, challenge a false premise, ask the user for clarification, or give up when resources are exhausted. While agent memory is often used to improve task success, its role in regulating these complex, reliable behaviors for UAQs has remained unclear. This paper investigates whether existing memory methods actually help, why they help, and what specific types of stored experience are most valuable.

Methods

The authors establish a unified agentic Retrieval-Augmented Generation (RAG) framework using a fixed Wikipedia-based interaction environment (utilizing Search and Lookup tools) with a maximum of 10 interaction rounds.

Figure 1: Overview of agent with external memory. Retrieved memory is injected as initial guidance before interaction begins. The resulting interaction trajectory is then distilled into new memory and stored for future reuse.

To understand how memory works, they operationally decompose its influence into two functional channels:

  1. Decision Guidance (G): Stored memory is injected into the initial context to directly shape the agent's high-level behavioral policy.
  2. Trajectory Shaping (T): Memory alters the step-by-step interaction trajectories generated during environment exploration.

They evaluate four representative memory baselines (Expel, MemEvolve, AWM, and AgentKB) and perform controlled content-level ablations across three memory families: Descriptive (compact trajectories), Rule-based (insights and principles), and Procedural (success traces and workflows).

Datasets

The framework is evaluated using three UAQ-related datasets containing balanced mixtures of answerable questions (ABQs) and unanswerable questions (UAQs):

  • KUQ: A high-quality, manually crowdsourced dataset featuring diverse data types and six categories of UAQs.
  • UAQFact: A dataset constructed from a knowledge graph where all data is supported by factual triples, featuring three types of UAQs.
  • RefuNQ: A evaluation-only dataset built by replacing query target entities in natural questions with non-existent entities (contains only one synthetic UAQ type).

Key Insights and Ideas

  • Dataset Shift vs. Model Shift: Memory is highly sensitive to changes in the data distribution. Moving memory from one dataset to another (cross-dataset transfer) often degrades performance. However, moving memory from one base model to another (cross-model transfer) is surprisingly stable. This implies that shifts in answerability patterns pose a much greater challenge to memory reuse than changes in the underlying LLM.
  • Guidance Over Trajectories: Memory improves UAQ handling primarily through decision guidance (telling the model how to behave conceptually) rather than trajectory shaping (the exact sequence of tool actions).
  • Procedural Memory Wins: Abstracting past experiences into structured, procedural formats—like workflows and success traces—is significantly more effective than storing raw or compressed interaction histories (compact trajectories).
  • The Polarity of Experience: The usefulness of successful vs. failed experiences depends entirely on how they are represented. Trajectory-style memories benefit almost exclusively from successful examples, whereas rule-based memories (like "Principles") can successfully extract valuable cautionary guidance from failures.

Key Results

Performance is measured using a Joint Score (JS), which balances Accuracy (Acc) on answerable questions and an Acceptable Ratio (AR) (refusal or clarification rates) on UAQs.

  • Model Dependency: On DeepSeek-V3.2, memory methods improved the Joint Score by an average of +1.88 percentage points, driven by large gains in AR. However, on Qwen3-235B, memory actually decreased the average Joint Score by -0.90 points, often triggering excessive caution and harming accuracy on answerable questions.
  • Channel Analysis: As shown in the functional channel decomposition, the Guidance-only (G) setting consistently outperformed the Trajectory-only (T) setting across stable memory methods, proving that high-level policy guidance is the primary driver of UAQ performance gains.
Figure 2: JS score comparison for different functional channels of memory: decision guidance (G), trajectory shaping (T), and their combination (G&T). The hori- zontal line denotes the No-memory setting performance (No-mem). Detailed results are listed in Appendix B.3.
  • Workflow Dominance: In single-content type evaluations, Workflow memory achieved the highest JS on DeepSeek-V3.2 (53.37 on KUQ vs. 41.55 for No-memory). Pairwise combinations that included procedural workflows yielded the strongest overall results.

Limitations and Future Work

The authors acknowledge several constraints in their study:

  1. Specificity of Objective: The findings focus strictly on reliable UAQ handling and may not generalize to tasks with completely different optimization goals.
  2. Environment Simplicity: The Wikipedia-based ReAct environment only provides basic Search and Lookup tools, failing to capture heterogeneous tool use or long-horizon multi-source evidence aggregation found in richer environments (e.g., GAIA or Deep Research systems).
  3. The Accuracy-Caution Trade-off: While memory successfully increases the Acceptable Ratio (AR), it often does so at the expense of Accuracy (Acc) on answerable questions due to over-refusal. The authors suggest that future work should explore answerability-aware memory retrieval or routing to dynamically apply different guidance depending on the nature of the question.

Summary updated

Abstract

Reliable handling of unanswerable questions (UAQs) is critical for trustworthy LLM-based agents. Although memory is widely used in agent systems, its role in reliable UAQ handling remains unclear. We present a systematic study of agent memory for UAQ handling under a unified agentic RAG framework, evaluating four representative memory methods across three UAQ-related datasets and two base models. We find that memory can improve UAQ performance in some settings, but such gains are selective rather than universal and remain fragile under dataset shift. Interestingly, cross-model memory reuse is often more feasible than cross-dataset transfer, suggesting that shifts in answerability patterns pose a greater challenge to memory reuse than changes in the base model itself. We further find that UAQ gains are more strongly preserved through decision guidance than through trajectory shaping, and that memory effectiveness depends strongly on representation. In particular, procedural and rule-based memories often provide the most reliable support for UAQ handling, while memory composition is most effective when procedural guidance is combined with complementary behavioral signals. Overall, our findings suggest that reliable UAQ memory depends less on storing larger amounts of experience and more on preserving transferable behavioral guidance.
BibTeX
@misc{tan2026what,
  author = {Chuanyuan Tan and Junjie Yu and Yuxin Wang and Yining Zheng and Xipeng Qiu and Wenliang Chen},
  title = {{What Makes Agent Memory Useful for Reliable Unanswerable Question Handling?}},
  year = {2026},
  eprint = {2608.27924},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2608.27924},
}
Figure