ReactLines

Overview

ReactLines is a specification that defines a formal model for expressing, scoping, retrieving, and composing atomic behavioral policies for cognitive agents. Each ReactLine is a structured instruction with 11 fields encoding who should do what, when, why, and under what constraints.

The 11-Field Schema

#	Field	Required	Description
1	id	Auto	Short unique identifier
2	name_token	Yes	Kebab-case reference token
3	scope	Yes	Hierarchical visibility level
4	priority	Yes	Conflict resolution weight (1–1000)
5	environment	No	Operational environment descriptor
6	context	No	Situational activation context
7	condition	Yes	Trigger condition (natural language)
8	action	Yes	Prescribed action (imperative)
9	outcome	No	Desired result from beneficiary's perspective
10	constraints	No	Structured limitations
11	actors	Yes	Participants (executor + beneficiary required)

Six Scopes

Scope	Lifetime	Example
TURN	Single message	Greeting, farewell
SESSION	Entire conversation	Tone maintenance
FLOW	Multi-step process	Onboarding steps
AGENT	All sessions for an agent	Persona, capabilities
ORG	All agents in org	Brand voice, compliance
GLOBAL	Universal	Safety, legal, ethics

Scopes are totally ordered: TURN < SESSION < FLOW < AGENT < ORG < GLOBAL.

Abstract

LLM agents in production rely on monolithic system prompts for behavioral control — static, context-blind, non-composable, and tightly coupled to specific frameworks. This yields instruction bloat, attention dilution, and unpredictable behavior as rule sets grow.

ReactLines defines atomic procedural behavioral policies with an eleven-field typed schema and a six-level hierarchical scoping system with priority-based conflict resolution. Rules are retrieved at runtime via embedding similarity, applying the RAG paradigm to procedural memory rather than factual knowledge.

Contributions: (C1) formal specification with eleven-field typed schema; (C2) first hierarchical scoping system for agent behavioral rules; (C3) semantic retrieval protocol applying RAG to behavioral specifications; (C4) open-source reference implementation with empirical validation.

Citation

@software{pymut2026reactlines,
  title={ReactLines: Atomic Procedural Memory Units for Cognitive Agents},
  author={Timana, Joel and Munoz, Diana and Munoz, Alvaro},
  year={2026},
  url={https://github.com/Pymut/reactlines}
}