Reflexio

What Reflexio does

Reflexio sits beside an agent and receives its published conversations. It extracts user profiles and playbooks from them, evaluates whether tasks succeeded, and feeds the lessons back so the agent avoids repeating mistakes and reuses strategies that worked. User-specific learnings stay scoped to that user, while lessons that recur across users can be aggregated into shared playbooks that are approved before reuse. It can also compare expert-written ideal responses with agent responses to extract playbooks from the differences.

The README reports a benchmark on four of five real GDPVal knowledge-work tasks with a Hermes agent running minimax/MiniMax-M2.7: a median 81 percent fewer planning steps and 72 percent fewer tokens, measured against a warm baseline in which the same agent reruns the task after learning on its own. It offers an SDK, a CLI and integrations, and a managed hosted version at Reflexio.ai with a 30-day Pro trial.

Key features

• Extracts user profiles and behavioral playbooks from published agent conversations
• Success evaluation of agent runs and reuse of successful execution paths
• Cross-user playbook aggregation with an approval step
• Learns from expert-provided ideal responses
• Python SDK and CLI, with a hosted managed option at Reflexio.ai

When to use it

• Making a Hermes agent stop repeating a mistake a user has already corrected
• Building shared playbooks from patterns across many users of one agent
• Measuring how much a learning layer reduces planning steps and tokens on repeated tasks

Who it is for: Developers who run agents with real users and want corrections and successful strategies to persist between sessions.

How it fits with Hermes Agent

Reflexio is a general self-improvement layer, and its published GDPVal benchmark was run on a Hermes agent.

Requirements: Python 3.12 or newer and uv for PyPI and source installs.

FAQ