PSYCHE: Defining Reproducible AI Agent Personalities with a Six-Layer Structure

PSYCHE is an open-source YAML-based AI personality configuration specification from PSM Labs. It uses six structured layers (Core, Persona, Shadow, Relations, Arc, Anchor) to address issues like personality drift and model migration failure in traditional system prompts, enabling consistent personality expression across models.

When building AI agents today, developers rely on natural language system prompts to define personalities, but there are fundamental flaws: when only one trait is defined, the model automatically fills in the rest; personality drifts gradually during conversations; safety boundaries become blurred after 10 rounds of dialogue; personality completely fails when switching models. PSYCHE attempts to solve these problems with structured configurations.

PSYCHE decomposes personality into six layers (with clear conflict resolution priorities):

1. Core Layer: Defines identity basics (name, type, origin, purpose) and the "fidelity" parameter, serving as the foundation of the personality;
2. Persona Layer: Quantifies traits (honesty, cooperativeness, etc.) using 10 floating-point values between 0.0 and 1.0, and defines voice style and anti-fluff rules;
3. Shadow Layer: Controls hidden content (secret goals, deception ability, etc.), which traditional prompts do not explicitly handle;
4. Relations Layer: Defines interaction styles with different objects (flattery level, anti-manipulation ability, etc.);
5. Arc Layer: Controls personality evolution over time (growth trajectory, drift monitoring, automatic correction);
6. Anchor Layer: Highest-priority absolute constraints (safety boundaries, emergency control, etc.), which take precedence in case of conflicts.

PSYCHE's core principle is "Define it, or the model will define it for you", using quantitative parameters to eliminate natural language ambiguity. For example, instead of natural language descriptions like "friendly but professional", it uses YAML to quantify traits (e.g., empathy:0.3, formality:0.4). Additionally, it has a built-in anti-fluff mechanism that allows enabling preset filters or customizing forbidden patterns (such as banning stylized expressions like "I'd be happy to"), making the personality more distinct and authentic.

PSYCHE provides an npm package @psmlabs/psyche that supports parsing, validation, and generation of system prompts; configurations can be written in YAML (suitable for integration) or converted to Markdown (suitable for direct pasting). The project claims compatibility with models like Claude, GPT-4, Gemini, LLaMA, and agent frameworks like OpenClaw, LangChain, AutoGPT.

Although PSYCHE's quantitative method solves the drift problem, it may struggle to accurately express subtle personality traits; whether continuous values can capture complex human personalities remains an open question. Additionally, the configuration of "deception ability" in the Shadow Layer raises ethical discussions: Is explicitly defining when to hide information safer or more dangerous than implicit handling?

PSYCHE represents a shift from "prompt engineering" to "personality engineering", replacing natural language descriptions with structured configurations and implicit inferences with explicit parameters, providing a feasible path for building reproducible, transferable, and auditable AI agent personalities. For production scenarios that require strict control over agent behavior, this specification is worth considering.