experience representation

Current large language model (LLM) agents face challenges in real-world deployment, including inefficiency, error-proneness, and poor maintainability, largely due to their reliance on on-the-fly reasoning and low-level tool invocation. This work introduces, for the first time, a skill-centric agent architecture that formalizes a comprehensive skill lifecycle framework encompassing representation, acquisition, retrieval, and evolution. It positions skills as a complementary mechanism bridging high-level reasoning and operational execution. By integrating key techniques—such as skill representation learning, automated acquisition, semantic retrieval, and continual evolution—and synergizing them with tool use, memory mechanisms, and contextual constraints, the proposed framework establishes a reusable and composable skill system. The paper further surveys representative approaches, open-source resources, and application scenarios, offering both theoretical foundations and practical guidance to enhance the scalability, robustness, and maintainability of intelligent agent systems.

In multi-turn tool-use scenarios, agents face dual challenges: progressive intent clarification and dynamic environmental evolution; existing approaches—relying rigidly on complete trajectories or fixed subtasks—struggle to adapt to temporal changes in state and information. This paper proposes the State-Integrated Tool Graph (SIT-Graph), the first framework unifying episodic memory (compressed state summaries) and procedural memory (tool dependency structures) to construct a dynamically retrievable, state-aware tool graph. During inference, SIT-Graph integrates historical sequence modeling, retrieval augmentation, and high-confidence path matching to enable context-driven, recall-execution co-decision making. Evaluated on multiple stateful, multi-turn tool-use benchmarks, it significantly outperforms strong baselines, improving both tool-selection robustness and cross-task experience transfer efficiency.

Existing memory systems for language agents address memory management: how to retrieve and page more information within a context budget. We address a complementary problem -- memory utility: what experience is worth keeping, and how it should change agent behavior. We present Atlas, a memory kernel that compiles accumulated task experience into an agent's instruction structure -- without fine-tuning, RAG, or human intervention. Memory is distillation, not storage; delivery is instruction rewriting, not context injection. Facts extracted from agent failures and successes are verified through a three-step promotion gate and delivered by rewriting the agent's system prompt with learned sub-bullets. On CUAD contract analysis, the evolved prompt improves GPT-4o token-level F1 by $+8.7$pp and precision by $+12.5$pp. On HotpotQA multi-hop QA, joint F1 improves $+3.16$pp. An ablation isolates the mechanism's defining property -- the training signal constraint: the evolved prompt learns exactly what it is taught, and nothing more. Applied to Claude Sonnet~4.5 using the same evolved prompt -- compiled from GPT-4o errors, unchanged -- joint F1 improves $+2.31$pp, with gains concentrating where Claude's stronger baseline leaves the most room -- confirming that the compiled knowledge is task-shaped, not model-shaped.

Large language model (LLM) agents face dual challenges in adapting to novel tasks: insufficient tool generalization and difficulty reusing prior experience. To address these, we propose SMITH, a cognitive memory architecture featuring a novel tri-level memory system—procedural, semantic, and episodic—that operates synergistically. SMITH enables adaptive toolset expansion via dynamic tool generation; supports cross-task experience reuse through episodic memory retrieval augmented by semantic matching; and enhances training efficiency via agent-aware difficulty re-estimation and curriculum learning. The method integrates hierarchical memory modeling, sandbox-driven iterative code generation, and semantic-similarity-guided memory retrieval. Evaluated on the GAIA benchmark, SMITH achieves 81.8% Pass@1 accuracy—significantly surpassing Alita (75.2%) and Memento (70.9%)—and provides the first empirical validation of continual capability evolution in LLM agents.

Automatically constructing high-quality, reusable skills from heterogeneous, fragmented interaction traces—often missing critical security behaviors—is highly challenging. This work proposes the W2S framework, which introduces a novel intermediate representation called RWSA to decouple skills into workflow structure, execution semantics, and runtime attachments, thereby enabling task decomposition, control-flow modeling, verification, rollback, and state management. W2S achieves efficient skill construction through trajectory segmentation, local skill draft generation, structural alignment, branch fusion, redundancy compression, and confidence-aware retention. Experimental evaluation across 70 skills demonstrates that W2S improves behavioral replay consistency by 10.5% compared to baseline approaches based on summarization and prompting.

This work addresses the limitations of large language model (LLM) agents in tool invocation, which are highly dependent on the quality of human-written tool interface descriptions and suffer significant performance degradation in cold-start scenarios with numerous candidate tools or absent execution traces. To overcome this, the authors propose Trace-Free+, a framework that leverages curriculum learning to transfer supervised knowledge from trace-rich environments to trace-free deployment settings. This approach guides the model to learn reusable tool-use patterns and automatically refine tool descriptions without relying on execution trajectories. Trace-Free+ supports cross-tool generalization and scales effectively to tool sets comprising hundreds of functions. Experiments on StableToolBench and RestBench demonstrate that Trace-Free+ substantially improves invocation accuracy on unseen tools, exhibiting strong cross-domain generalization and robustness at scale.

This work addresses the challenge that agents struggle to efficiently reuse successful experiences when repeatedly performing similar tasks, often resulting in redundant reasoning and excessive interaction rounds. To overcome this limitation, the paper introduces a novel framework that formalizes procedural skills as parameterized finite state machine (PFSM) subgraphs and automatically extracts, verifies, and reuses structured skills through distillation and compilation of successful execution trajectories. Evaluated on the ALFWorld and WebArena benchmarks, the proposed method significantly improves task success rates while reducing the number of required interactions, demonstrating effectiveness across language models of varying scales.

This work addresses the challenge of routing errors in enterprise AI agent systems caused by overlapping skill descriptions, a problem traditionally mitigated through costly manual tuning. The authors propose a lightweight, automated approach that requires only a single large language model (LLM) rewrite of skill descriptions, supplemented by a small set of misrouted examples, to substantially improve routing accuracy. They introduce the training-validation F1 gap as a diagnostic metric for architectural intervention and establish a systematic evaluation framework. Evaluated on a production system with nine skills, the method achieves an average F1 score of 79.2%, matching the performance of manual tuning while reducing per-skill optimization time from 120 minutes to 3.8 minutes—a 32-fold efficiency gain. These results demonstrate that complex design yields diminishing returns, whereas a simple, streamlined solution is both effective and highly efficient.

This work proposes the MSCE framework, which introduces the first training-free mechanism for the co-evolution of memory and skills in long-horizon LLM agents. Unlike existing memory systems that treat historical trajectories as passive context, MSCE actively structures experience into concrete step-wise trajectories, reusable procedural policies, and environmental knowledge. It dynamically distills, validates, and retrieves reliable skills through evidence-anchored L2 policy crystallization and reflection-weighted value backfilling. The framework enables precise skill boundary delineation, decision guidance, and reliability estimation. Evaluated on EvoAgentBench and LoCoMo benchmarks, MSCE significantly outperforms current skill-augmented and memory-driven approaches, demonstrating exceptional cross-domain transfer and continual evolution capabilities.

Existing skill rewriting approaches often reduce the problem to prompt compression, overlooking their impact on agent execution cost and task quality. This work reframes skill design as cost-aware operational knowledge engineering, introducing a controllable framework to analyze skill structure and perform rewriting based on information retention strategies. Using the SkillsBench benchmark, the study systematically evaluates diverse anchoring strategies—such as API/code, workflows, and rules/formulas—under fixed tasks, environments, and validators, thereby refuting the existence of a universal template. Experimental results demonstrate that the proposed method reduces total cost by 7.0% and downstream token cost by 6.0% in primary evaluation; under frozen cross-model transfer, average reductions reach 14.7% and 13.7%, respectively, while preserving validation quality.

This work addresses the limitations of existing large language model agents in tool use, which rely on agent-side optimization and thus suffer from non-shareable tool knowledge constrained by historical tasks. The authors propose ToolAtlas, a graph-based framework that, for the first time, enables persistent storage of reusable tool knowledge—including capabilities, failure boundaries, and cross-tool compositions—at the tool provider side. Agents efficiently query this knowledge through capability-guided adaptive graph traversal. ToolAtlas facilitates cross-environment and cross-framework knowledge transfer without requiring retraining or online exploration. Evaluated on two MCP benchmarks covering eight services, ToolAtlas achieves performance gains of up to 21.61% and 18.61% in pass@1 and pass@4 metrics, respectively, with relative improvements of 24.16%/16.22% and 17.49%/14.27% in cross-environment and cross-framework transfer settings.