duyvuleo/awesome-agents-that-learn

A research guide to self-learning, self-improving, and self-evolving AI agents.
How do agents turn experience into something useful on the next task? This collection follows the research from reflection and reusable memories to learned skills, agent redesign, self-play, and model updates. Each work explains what changes, what drives the change, and what evidence supports improvement.
Explore the collection · Start reading · Browse papers · Compare systems · Compare benchmarks
109 research works (94 papers, 6 surveys, 9 benchmark papers) · 23 systems and frameworks · 8 reading routes
Source review: 2026-10-29. For research works, 49 entries include inspection of selected full-text sections and 60 are based on primary abstracts/metadata. Systems are separate project-documentation/source inspections. No linked research code was executed and no published result was independently reproduced for this release.
## Start here
1. Read the short primer to separate within-task correction from persistent learning.
2. Choose a reading route and compare mechanisms, rather than reading by date alone.
3. Use the evaluation guide to ask whether gains survive new tasks, a fair baseline, and a fixed budget.
4. Keep your own notes with the study template.
5. Compare systems and frameworks by capabilities, requirements, license, and paper relationship before choosing an implementation to study.
## Choose a reading route
| Your question | Route |
| What makes an agent a learner? | Foundations and boundaries |
| How does experience become reusable knowledge or a skill? | Memory and skills |
| Can an agent’s prompts and workflow improve? | Agent design and optimization |
| Where do self-generated training signals come from? | Learning in model weights |
| Can agents generate their own curriculum? | Self-play and curricula |
| What does it mean to improve the improver? | Recursive improvement |
| How can we tell whether learning actually helped? | Evaluation and failure modes |
| Does learning persist without forgetting or drift? | Continual learning |
| What software can I inspect or build on? | Systems and frameworks, with availability and version notes |
## A few entry points
These are complementary starting points, not a leaderboard. Their tasks, feedback, and budgets differ.
| Work | Learn this idea |
| Voyager: An Open-Ended Embodied Agent with Large Language Models | Builds an expanding library of executable Minecraft skills, retrieved and composed for later tasks. An automatic curriculum proposes objectives, while execution feedback and a model-based critic guide program repair before a skill is retained. |
| ExpeL: LLM Agents Are Experiential Learners | Collects training-task trajectories, contrasts successes and failures to extract reusable textual insights, and retrieves relevant experiences for new tasks. Learning changes the external experience and insight stores without updating the base model. |
| DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines | Represents language-model pipelines as declarative modules and compiles demonstrations for their component prompts. The reusable artifact is an optimized pipeline configuration. |