[CVPR 2026] Memento-Skills: The Rise of the Agent-Designing Agent

TL;DR

Memento-Skills is a breakthrough in continual learning for LLMs that side-steps the need for expensive fine-tuning. Instead of re-training the model, it treats "Skills" (structured markdown, code, and prompts) as a mutable, evolving memory. By implementing a Read-Write Reflective Learning loop, the system allows a frozen LLM to "design" better versions of itself, achieving a relative improvement of up to 116% on expert-level benchmarks.

Motivation: The "Stateless Agent" Problem

Most modern agents are deployed with fixed parameters ($heta$). When they fail, they stay failed. The authors argue that while standard LLMs are "frozen," their behavior shouldn't be.

The core bottleneck in current RAG-based agents is that semantic similarity $eq$ behavioral utility. Just because a past case looks like the current problem doesn't mean the solution is the right tool to execute. Memento-Skills shifts the paradigm from "remembering what happened" to "refining how to act."

Methodology: The Read-Write Evolution Loop

The architecture is governed by the Stateful Reflective Decision Process (SRDP). It approximates policy iteration through two core phases:

1. Read (Policy Improvement): A behavior-aligned router (trained via offline RL) selects the most execution-relevant skill from an 8k+ skill library.
2. Write (Policy Evaluation & Mutation): If a task fails, a "Failure Attribution" module identifies the responsible skill. A "Skill Rewriter" then patches the code or prompt, followed by a Unit Test Gate to ensure no regressions.

Figure: The Read-Write Reflective Learning loop. The LLM remains frozen; all adaptation occurs in the external skill memory.

The Secret Sauce: Behavior-Aligned Routing

Unlike standard RAG that uses cosine similarity, Memento-Skills uses InfoNCE loss to train its router. By treating retrieval as a "one-step MDP," the router learns to predict which skill will actually lead to a successful "Judge" score, essentially fitting a Q-function over the library.

Experiments: Breaking the Limits of "Humanity's Last Exam"

The system was tested on GAIA (General AI Assistants) and HLE (Humanity’s Last Exam).

• HLE Results: On this expert-level academic benchmark, the self-evolving mechanism allowed the agent to transfer knowledge across subjects (e.g., Biology, Humanities). The success rate climbed from 17.9% to 38.7% on the test set.
• Convergence: The authors show that as the skill library grows and densifies, the "memory coverage radius" ($r_M$) shrinks, leading to a mathematically provable reduction in the performance gap.

Figure: HLE test set comparison showing Memento-Skills more than doubling the baseline accuracy.

Deep Insight: Skills as Muscle Memory

The most striking visualization in the paper is the t-SNE projection of skill embeddings. Learning on HLE expanded the library from 5 "atomic" seed skills to 235 specialized clusters.

Figure: Evolutionary densification of the skill space. Blue dots represent newly "discovered" skills.

This process mimics biological motor learning: what starts as deliberate, high-level reasoning eventually consolidates into robust, "compiled" routines (skills).

Conclusion & Future Outlook

Memento-Skills proves that intelligence can be externalized. By modularizing an agent's capabilities into a searchable, writable library, we can achieve perpetual self-improvement without a single gradient update to the foundation model.

Limitations: The system currently relies on an LLM-as-a-Judge for rewards, which can be a source of bias. Furthermore, as the library hits millions of skills, the latency of retrieval and the complexity of "skill deduplication" will become significant engineering hurdles.