Reflexion

Gold definitionUpdated Apr 2, 2026

Reflexion is an advanced framework for building autonomous agents powered by Large Language Models (LLMs), designed to overcome the limitations of single-pass reasoning. At its core, Reflexion allows an LLM agent to observe its own actions and the resulting outcomes, then generate reflective feedback based on these experiences. This feedback, stored in a memory buffer, is subsequently used to inform and improve the agent's future decision-making and planning processes. The mechanism involves an iterative loop where the agent attempts a task, evaluates its success or failure, reflects on 'why' it succeeded or failed, and then incorporates these insights to guide its next attempt or subsequent steps. This approach significantly enhances the agent's robustness, enables error recovery, and improves its ability to tackle complex, multi-step problems that require sustained reasoning and adaptation. Reflexion is primarily used by researchers in AI agents, reinforcement learning, cognitive architectures, and autonomous systems development, aiming to create more capable and reliable intelligent agents.

Core Mechanism of Reflexion

Iterative Self-Reflection: The agent executes a sequence of actions, observes the environment, and then prompts an LLM to generate a 'reflection' on the success or failure of its trajectory. This reflection identifies critical mistakes or successful strategies, providing actionable insights for future steps.
Memory and Experience Replay: Past trajectories, including actions, observations, and generated reflections, are stored in a memory buffer. When facing a new or similar problem, the agent can retrieve relevant reflections to guide its planning and action generation, learning from its own history.
LLM as the Reflective Engine: A key aspect of Reflexion is leveraging the LLM's advanced reasoning capabilities not just for task execution, but also for meta-cognition. The LLM analyzes its own performance, identifies patterns, and synthesizes high-level advice, effectively acting as its own critic and mentor.

Benefits and Applications of Reflexion

Enhanced Problem Solving: Reflexion significantly improves an agent's ability to solve complex, multi-step tasks that require planning and error correction. By iteratively refining its approach, the agent can navigate intricate environments and achieve goals that are challenging for non-reflective systems.
Robustness and Error Recovery: The framework allows agents to identify and recover from mistakes made during execution. Instead of failing outright, the agent can reflect on the error, understand its cause, and adjust its strategy to avoid similar pitfalls in subsequent attempts or future tasks.
Task Generalization: By generating and storing abstract reflections, agents can learn generalizable strategies from specific task experiences. These learned insights can then be applied to novel, but related, problems, fostering a form of meta-learning or transfer learning.

Challenges and Future Directions for Reflexion

Computational Cost: Generating reflections and re-planning often involves multiple calls to a large language model, which can be computationally expensive and slow. Optimizing the reflection process and reducing LLM inference costs are ongoing research challenges.
Quality of Reflection: The effectiveness of Reflexion heavily depends on the quality and accuracy of the LLM-generated reflections. Ensuring the LLM produces insightful, non-hallucinated, and actionable feedback is crucial, especially in complex or ambiguous scenarios.
Integration with Planning: Future work involves deeper integration of reflection with advanced planning algorithms, allowing reflections to not just guide re-planning but also to refine the planning process itself, leading to more sophisticated and efficient agent behaviors.

At a glance

Executive summary

Reflexion is a method for AI agents to learn from their own experiences by reflecting on past actions and outcomes. This self-correction process helps them improve their performance on difficult tasks, making them more robust and capable of solving complex problems over time.

TL;DR

Reflexion lets AI agents learn from their mistakes by thinking about what went wrong or right, then using that knowledge to do better next time.

Key points

LLM agents generate and utilize self-reflections on past actions and outcomes.
Solves the problem of improving LLM agent performance on complex, multi-step tasks requiring iterative refinement.
Used by AI agent researchers, developers of autonomous systems, and those exploring cognitive architectures for LLMs.
Unlike single-pass prompting, Reflexion enables iterative self-correction and learning from experience.
Represents a key research trend towards more autonomous, self-improving, and robust LLM-based agents.

Use cases

Automated code debugging: An agent writes code, runs tests, reflects on failures, and iteratively refines the code.
Complex robotic task execution: A robot plans actions, observes results, reflects on discrepancies, and adjusts its physical movements.
Interactive story generation: An agent generates story segments, evaluates coherence, reflects on plot holes, and revises the narrative.
Scientific experiment design: An agent proposes experiments, analyzes simulated results, reflects on hypotheses, and refines the experimental setup.

Also known as

Self-Refine, Self-Correction, Self-Improvement, Cognitive Architectures for LLMs