Memento-Skills: Evolving Agent Behaviors

Updated 3 July 2026

Memento-Skills are explicit, persistent agent routines encoded in structured artifacts that enable continual learning without parameter updates.
They employ embedding-based retrieval and read–write reflective learning to select, refine, and evolve task-specific procedures.
Empirical benchmarks show substantial performance gains, highlighting robust compositional reasoning and adaptive skill evolution.

Memento-Skills denote a class of explicit, persistent, and generalizable agent skills, typically represented as structured artifacts (e.g., code snippets, prompts, markdown folders, memory routines) that encode behaviors, heuristics, strategies, or task-specific procedures. These skills provide a substrate for continual learning, compositional reasoning, and autonomous agent evolution within both LLM-centric and broader neuro-symbolic frameworks. Memento-Skills have emerged as a response to inherent limitations of static or prompt-engineered models, enabling agents to adapt and improve not through parameter updates but via the evolution and organization of externalizable, reusable skills.

1. Formal Definition and Conceptual Basis

Memento-Skills are defined as structured, explicit routines, policies, or artifacts that an agent can read, write, retrieve, refine, and compose to solve a wide range of tasks. These entities persist across interactions, function as both memory and procedure, and enable continual agent improvement without parameter finetuning. In the foundational work introducing the Memento-Skills formalism, skills are realized as directories containing markdown specs, prompt templates, helper code, and unit tests; these persist as the agent’s external, modifiable “skill memory” (Zhou et al., 19 Mar 2026).

The conceptual basis stems from two streams:

Cognitive inspiration: The notion of “metamemory” in cognitive science encapsulates learned competencies to encode, retrieve, and organize knowledge. In LLM agents, this translates to models manipulating memory explicitly (e.g., via file-system–like memory actions) and treating memory management as a trainable, observable skill (Wu et al., 1 Jul 2026).
Agent skill formalism: Memento-Skills are units of behavior that encode both context and action, in contrast to prompt-based heuristics or parameter frozen policies. They generalize standard skills by making the skill acquisition, retrieval, and revision process persistent and composable across episodes (Zhou et al., 19 Mar 2026, Zhang et al., 2 Feb 2026).

2. Memory-Augmented Reinforcement and Read–Write Reflective Learning

Memento-Skills operationalize a memory-based reinforcement learning framework in which skills both inform the agent’s state (via stateful prompts or external memory) and serve as actions (retrieval and execution). States comprise tasks and memory, while actions include invoking or editing skills. The agent iteratively selects relevant skills based on contextual embeddings and applies them to generate action trajectories (Zhou et al., 19 Mar 2026). Rewards from downstream execution, as judged by automated or human-in-the-loop supervisors, drive the evolution of the skill library via a closed-loop Read–Write Reflective Learning mechanism. This mechanism decomposes each episode into “read” (skill retrieval/selection using an embedding-aligned router) and “write” (skill refinement, correction, or discovery based on outcomes and correction signals).

Formally, agent transitions are defined over joint state-memory-action tuples, e.g.,

$T^\pi(a\mid s, M_t) = \sum_{c\in M_t}p(c\mid s,M_t)\;P_{\mathrm{LLM}}(a\mid s,c)$

and memory is updated as

$M_{t+1} = \mathrm{Write}(M_t, s_t, a_t, r_t)$

under the Reflected-MDP formalism (Zhou et al., 19 Mar 2026).

3. Skill Representation, Organization, and Retrieval

Each Memento-Skill is typically realized as a structured directory (“skill folder”) containing:

Declarative specification (SKILL.md): a description of inputs, outputs, preconditions, and usage examples.
Prompt template (prompt.tpl): templates for grounding skills at inference.
Executable code (main.py): helper functions or automated procedures.
Unit tests (unit_test.py): correctness checks.

Skill selection is formulated as a retrieval problem over embedding spaces. A behavior-trained router ranks the available skills using query- and skill-specific embeddings via score functions such as:

$s(d,q) = e(d)^\top u(q)$

followed by retrieval policies optimized for expected utility/reward (Zhou et al., 19 Mar 2026, Zhang et al., 2 Feb 2026). “Tip memory” (recent natural language correction hints) may be interleaved into prompts for rapid adaption.

Skills can be created de novo (when the router fails to match an existing skill) or optimized in situ via LLM-guided mutation (prompt updates, new test cases, code corrections) with all changes guarded by automated regression checks.

4. Continual, Parameter-Free Learning and the Agent-Designing Agent

A central property of Memento-Skills is that all adaptation and improvement is realized through modification of the external skill library, not through finetuning or gradient updates to the LLM backbone parameters. This design enables:

Provable convergence: Under mild assumptions on reward boundedness and coverage, iterative read–write learning improves expected returns as the memory becomes richer and the skill router better optimized (Zhou et al., 19 Mar 2026).
Zero-retraining continual learning: Agents evolve capabilities via skill mutation/discovery; no model gradient descent is required at deployment.
Meta-agent capabilities: The system can act as an “agent-designing agent,” autonomously synthesizing new task-specific agents by composing or specializing existing skills. For unseen tasks, the agent instantiates new skill folders with spec, prompt, code, and unit test, then iteratively refines through further experience and reflect–write cycles (Zhou et al., 19 Mar 2026).

5. Empirical Performance and Benchmark Results

Memento-Skills have demonstrated strong empirical improvements on generalist, hard-to-script benchmarks. On the General AI Assistants (GAIA) benchmark, the system reached 66.0% test accuracy—26.2% relative improvement over a baseline excluding the skill evolution mechanism. On the Humanity’s Last Exam (HLE), a challenging compositional benchmark, Memento-Skills achieved 38.7% test accuracy (116.2% relative improvement over baseline) (Zhou et al., 19 Mar 2026). Skill library size grew commensurately with coverage (e.g., 41 new skills in GAIA, 235 in HLE, clustering by domain, Fig. 12 in (Zhou et al., 19 Mar 2026)).

Router ablations confirm the necessity of high-quality retrieval: Recall@1 rose from 0.32 (BM25) / 0.54 (Qwen3) to 0.60 (Memento-Qwen), with a judge success rate of 0.80. All adaptation occurs external to core model weights, highlighting the viability of library-driven agent evolution.

The Memento-Skills approach distinguishes itself from related techniques by the explicit, modular, persistent, and revisable nature of its skills:

Contrast with prompt-hint internalization: Context distillation methods (“Memento No More”) transfer ephemeral prompt guidance into model parameters, reducing prompt bloat but requiring supervised gradient updates (Alakuijala et al., 3 Feb 2025). Memento-Skills retain adaptation external to the model.
Skill banks and procedural memories: Works such as MemSkill formalize skill banks for memory management; these evolve via analysis of hard cases and LLM-guided proposal of new skills. Both skill selection and skill evolution are policy-learnable (Zhang et al., 2 Feb 2026).
Integration with memory management: AutoMem supports memory management as a trainable skill but focuses on LLM proficiency and memory schema optimization rather than agent-designed procedural skills (Wu et al., 1 Jul 2026).
Comparison with procedural/progressive learning: Agents such as MEMO, which aggregate local feedback into general, parameterized capabilities, provide mechanisms for emergent Memento-Skills via trajectory clustering and code-template retrofitting (Christie et al., 4 Mar 2026).

7. Limitations, Open Challenges, and Future Directions

Memento-Skills scale with the richness of the task and domain, but several practical limitations and research challenges persist:

As the skill library grows, retrieval latency and skill redundancy may become bottlenecks. Embedding spaces can suffer diminishing returns as coverage saturates.
Executing synthesized skills requires robust sandboxing, automated verification, and out-of-distribution handling to ensure reliable and safe agent behavior.
Quantitative rates of skill library convergence and transfer remain open for analysis.
Future investigations may integrate formal verification and enhanced test harnesses, explore multi-agent skill sharing at scale, and develop hybrid parametric/non-parametric schemes for high-value skill adaptation (Zhou et al., 19 Mar 2026).

Memento-Skills collectively establish an architectural and algorithmic paradigm for agentic systems: adaptation and generalization are externalized as explicit, evolvable skills, enabling LLM-based agents to design, refine, and construct new behaviors on a continual, modular, and parameter-invariant basis.