Memento-Skills: Let Agents Design Agents

Abstract: We introduce \emph{Memento-Skills}, a generalist, continually-learnable LLM agent system that functions as an \emph{agent-designing agent}: it autonomously constructs, adapts, and improves task-specific agents through experience. The system is built on a memory-based reinforcement learning framework with \emph{stateful prompts}, where reusable skills (stored as structured markdown files) serve as persistent, evolving memory. These skills encode both behaviour and context, enabling the agent to carry forward knowledge across interactions. Starting from simple elementary skills (like Web search and terminal operations), the agent continually improves via the \emph{Read--Write Reflective Learning} mechanism introduced in \emph{Memento~2}~\cite{wang2025memento2}. In the \emph{read} phase, a behaviour-trainable skill router selects the most relevant skill conditioned on the current stateful prompt; in the \emph{write} phase, the agent updates and expands its skill library based on new experience. This closed-loop design enables \emph{continual learning without updating LLM parameters}, as all adaptation is realised through the evolution of externalised skills and prompts. Unlike prior approaches that rely on human-designed agents, Memento-Skills enables a generalist agent to \emph{design agents end-to-end} for new tasks. Through iterative skill generation and refinement, the system progressively improves its own capabilities. Experiments on the \emph{General AI Assistants} benchmark and \emph{Humanity's Last Exam} demonstrate sustained gains, achieving 26.2\% and 116.2\% relative improvements in overall accuracy, respectively. Code is available at https://github.com/Memento-Teams/Memento-Skills.

• The paper introduces Memento-Skills, a system enabling parameter-free continual learning by evolving external skill memories rather than updating LLM weights.
• It employs a reflective, contrastive RL-trained skill router to dynamically retrieve, execute, and improve modular skill artefacts for adaptive task handling.
• Empirical results on GAIA and HLE benchmarks demonstrate significant improvements in test accuracy and effective domain transfer via progressive skill evolution.

Memento-Skills: Autonomous Skill-Driven Continual Learning for LLM Agents

Introduction and Motivation

The "Memento-Skills" system (2603.18743) operationalizes a memory-based continual learning framework for LLM agents via explicit, self-evolving skill repositories. Departing from the paradigm of static, frozen LLM agents—or expensive, parameter-updated pipelines—Memento-Skills proposes a deployment-centric approach where experience-driven adaptation occurs through structured, executable skill artefacts rather than gradient-based weight updates. This shifts the locus of learning from the internal parameters of the LLM to an externalized, modular skill memory, transforming agents from one-shot automatons into entities capable of persistent, progressive self-improvement without retraining.

Theoretical Foundation: Skill-Level Reflective Memory

Memento-Skills is grounded in the "Stateful Reflective Decision Process" (SRDP) formalism, which extends the traditional MDP by incorporating an episodic memory $M_t$ of structured, declarative skills. Each skill encapsulates not only action procedures but also contextual and diagnostic information, with the memory evolving as the agent interacts with tasks. Formally, the agent’s policy $\pi(a|s, M_t)$ is conditioned jointly on the state and the current skill memory—preserving the Markov property.

The closed-loop dynamics of "Read-Write Reflective Learning" are central: the agent (1) observes the task, (2) retrieves the most relevant skill via a behaviour-aligned router, (3) executes the proposed plan, (4) incorporates post hoc feedback, and (5) mutates or invents skills to address failures. Policy improvement and evaluation thus manifest directly at the skill level, enabling continual learning without parameter tuning of the core LLM. Convergence guarantees are formally established (see Theorem 1.3 and relevant extensions in [17]), confirming that under bounded rewards and discounting, the system’s routing policy and skill set approach optimality.

Architecture and Implementation

Memento-Skills encapsulates the agent in a modular orchestration architecture:

• Skill Memory: Structured as directories containing code, prompts, and specifications (e.g., SKILL.md). Skills are first-class, independently improvable memory units.
• Skill Router: Behaviours are routed using a contrastive, one-step RL-trained retrieval engine. Unlike classical BM25 or dense embedding search, the router is optimized for action success (behavioural utility) rather than textual similarity, leveraging synthetic negative samples and InfoNCE-based objectives.
• Skill Evolution Engine: After each task, failures are attributed to the responsible skill by an LLM-based judge; these skills are then selectively rewritten or, if generalization fails, replaced via discovery of new skills. Automatic unit tests constrain regressions.
• Infrastructure: All agent interaction, execution, and memory management is compartmentalized from LLM inference, supporting separation of concerns and robust deployment.

Tasks are received via CLI/GUI, then decomposed and routed to the best-suited skill or synthesized anew. Experience and diagnostic feedback iteratively enhance both retrieval and execution competence.

Empirical Evaluation

Benchmarks

Two principal benchmarks are used:

• General AI Assistants (GAIA): Requires heterogeneous, multi-step reasoning and real-world tool use.
• Humanity’s Last Exam (HLE): A broad-domain, subject-diverse academic challenge measuring transfer and abstraction.

Results

• GAIA: Memento-Skills achieves a 66.0% test accuracy, compared to 52.3% for the static Read-Write ablation—a 13.7 percentage-point improvement with three rounds of reflective training.
• HLE: The system reaches 38.7% test accuracy (vs. 17.9% for the baseline), with notable transfer in structured domains (e.g., 66.7% in Humanities, 60.7% in Biology). Relative improvements exceed 116% in some tracks, underscoring the effectiveness of reflective skill evolution, especially when domain-aligned clusters enhance transferability and memory coverage.
• The Memento-Qwen router outperforms BM25 and generic embeddings in both recall and end-to-end execution rates: Recall@1 is 0.60 (+10% over embedding baselines); judge success rate is 0.80.
• The skill library grows from 5 atomic skills to over 235 through interaction, with semantic clustering matching task taxonomy. Diminishing returns in accuracy are observed as the library covers the input space more densely, matching theoretical predictions for convergence and memory coverage shrinkage.

Claims and Implications

Key claims and technical advances:

• Parameter-free continual learning: All agent improvement is driven by executable skill adaptation, with no LLM weight updates—substantially reducing computational cost and deployment friction.
• Skill-based modular memory: Skills, not raw traces, serve as the fundamental memory artefact, enhancing traceability, modularity, and policy evolvability.
• Behaviour-optimized retrieval: Contrastive RL training of the router provides measurable gains in routing precision over semantic similarity-based methods.
• Empirical validation: Substantial, statistically robust improvements in accuracy demonstrate the importance of full reflective skill evolution (beyond naive archive/retrieval policies), with domain-aligned transfer as a critical enabler of agent generalization.

Theoretical and Practical Implications

This framework suggests that the scalability limitations of gradient-based continual learning for LLM agents can be circumvented by treating programmable skills as persistent external memory. Adaptive modular skill repositories decouple learning timescales from LLM parameterization, enabling persistent adaptation in compute- and data-limited regimes. As skills are independently improvable, the architecture supports compositionality, regression testing, and verifiable safety properties—features absent in monolithic LLM weights.

The approach also exposes distinct axes for further performance enhancement: upgrading the LLM or router embedding model, or simply running more episodes to densify memory coverage. Safety evaluations and failure analysis become tractable, as artefactual skills can be audited and isolated, in contrast to inscrutable weight perturbations.

Future Directions

The Memento-Skills framework opens several research trajectories:

• Scalable skill management: The evolution engine must address redundancy, abstraction, and compositional skill bootstrapping as libraries scale to millions of artefacts.
• Advanced routing and transfer: As skills proliferate, hierarchical and context-aware retrieval strategies will be needed to further minimize retrieval error and maximize transfer.
• Safety, verifiability, and control: The modular nature of skill artefacts enables fine-grained safety constraints and interpretable reasoning paths, a necessary foundation for regulated deployment of generalist AI agents.
• Integration with alternative LLM kernels: As LLM architectures advance, decoupling agent adaptation from LLM weights ensures compatibility and rapid deployment of new architectures.

Conclusion

Memento-Skills constitutes a substantial advance in continual, skill-based learning for LLM agents by shifting adaptation into a dynamically improving, externally managed skill library. This architecture achieves robust gains in agent proficiency and transfer capability across demanding benchmarks—without model retraining—by surfacing and optimizing skills as the central memory unit. This paradigm establishes a foundation for persistent, auditable, and modular intelligence, redefining how autonomous agents can adapt and evolve post-deployment.