ReflexioAI/reflexio

What it solves

Reflexio is an AI agent self-improvement harness designed to prevent agents from repeating mistakes and to lock in successful strategies. It solves the problem of agents starting from scratch in every session by turning real user interactions, corrections, and expert feedback into persisted behavioral improvements and reusable playbooks.

How it works

Reflexio closes the self-improvement loop by processing conversations published from an agent. It extracts stable facts about users into User Profiles and identifies behavioral patterns to create User Playbooks. Recurring patterns across different users are then aggregated into Agent Playbooks, which can be reviewed and approved for global use. Additionally, it can compare agent responses against human-expert "ideal responses" to generate actionable SOPs (Standard Operating Procedures).

Who it’s for

It is built for developers of AI agents who want their systems to become smarter and more efficient over time through continuous learning from real-world usage without requiring constant retraining of the underlying models.

Highlights

• Self-Correction: Transforms user corrections into improved decision-making to avoid repeating the same errors.
• Expert Learning: Automatically extracts actionable playbooks by comparing agent responses with expert-provided ideal responses.
• Playbook Aggregation: Clusters similar user-level learnings into shared agent playbooks that improve the agent for all users.
• High-Performance Retrieval: Features hybrid search (vector + full-text) across profiles and playbooks with low latency (~57ms p50).
• Broad LLM Support: Compatible with multiple providers including OpenAI, Anthropic, Google Gemini, and DeepSeek via LiteLLM.