Jungian Psyche

The psyche system models the agent's personality, ethical constraints, and behavioral tendencies using Carl Jung's analytical psychology. It replaces rigid rule-based ethics (Asimov's laws) with a dynamic, evolvable system that learns from experience.

The psyche is a special Core compartment that influences four aspects of the OODA loop:

1. Tool selection -- archetype weights bias which tools the agent prefers.
2. Action gating -- shadow veto patterns block dangerous actions.
3. Reflection -- the psyche evolves based on the agent's performance.
4. Output style -- the persona's grammar preference controls narrative synthesis.

Structure

Psyche
  ├── Persona           <- outward mask (communication style)
  ├── Shadow            <- constrained anti-patterns (vetoes + biases)
  ├── ArchetypeWeights  <- behavioral tendencies (tool selection bias)
  └── SelfIntegration   <- growth tracking (individuation)

Persona

The Persona controls how the agent communicates. It is the "mask" the agent presents to the user.

[persona]
name = "Scholar"
grammar_preference = "narrative"        # or "formal", "terse", or a custom TOML path
traits = ["precise", "curious", "thorough"]
tone = ["clear", "methodical"]

When the agent synthesizes output (Agent::synthesize_findings()), it checks the persona's grammar_preference and uses it as the default grammar archetype. The grammar framework then structures the output accordingly:

Shadow

The Shadow represents the agent's ethical constraints -- things it must not do (vetoes) and things it should avoid (biases).

[shadow]

[[shadow.veto_patterns]]
name = "destructive_action"
triggers = ["delete all", "drop table", "rm -rf"]
severity = 1.0
explanation = "Destructive actions require explicit user confirmation."

[[shadow.bias_patterns]]
name = "repetitive_loop"
triggers = ["same tool", "repeated"]
severity = 0.3
explanation = "Detected repetitive pattern - try a different approach."

Two severity levels

Trigger matching: Each trigger string is checked as a case-insensitive substring against the action description ("tool={name} input={:?}"). Multiple triggers in a pattern are OR'd -- any match fires the pattern.

When a veto fires (during the Act phase):

1. The action is blocked -- it never reaches the tool registry.
2. A DerivationKind::ShadowVeto provenance record is stored.
3. psyche.record_shadow_encounter() increments the shadow encounter counter.
4. The cycle returns a GoalProgress::Failed with the veto reason.
5. The agent adapts on the next cycle by choosing a different tool.

Archetypes

Four Jungian archetypes bias tool selection during the Decide phase:

Scoring formula: For each tool candidate, the archetype bonus is:

archetype_bonus = (archetype_weight - 0.5) * 0.15

Examples with default weights:

• kg_query (Sage, weight 0.7): bonus = +0.030
• gap_analysis (Healer, weight 0.5): bonus = 0.000
• memory_recall (Guardian, weight 0.4): bonus = -0.015

The bonus is added to the tool's total score alongside base_score, recency_penalty, novelty_bonus, episodic_bonus, and pressure_bonus. The effect is subtle per-cycle but cumulative over time.

Self-Integration (Individuation)

The Self is Jung's integrative center. In the agent, it tracks psychological growth through experience.

[self_integration]
individuation_level = 0.1     # [0.0, 1.0] - how integrated the psyche is
last_evolution_cycle = 0
shadow_encounters = 0         # times shadow patterns were triggered
rebalance_count = 0           # times archetypes were rebalanced
dominant_archetype = "sage"   # derived from highest archetype weight

Individuation growth formula (applied during reflection):

growth = 0.01 * min(shadow_encounters, 5)
individuation_level = min(individuation_level + growth, 1.0)

Encountering and acknowledging the shadow (having actions vetoed, having goals abandoned) drives psychological growth -- mirroring Jung's concept that integrating the shadow is the path to individuation.

Psyche Evolution

Evolution happens automatically during the agent's periodic reflection cycle. When reflect() is called, it passes the psyche to Psyche::evolve(), which:

1. Rebalances archetypes based on tool effectiveness:
   • Consistent success (>70% rate, 2+ uses): archetype weight +0.02
   • Flagged ineffective (<30% rate, 2+ uses): archetype weight -0.02
   • Weights clamped to [0.1, 0.95]
2. Acknowledges shadow encounters: Incrementing counter on goal abandonment.
3. Grows individuation: individuation_level += 0.01 * min(shadow_encounters, 5)
4. Updates dominant archetype: Recomputed from current weights.

OODA Loop Integration

The score breakdown is visible in the decision reasoning:

[score=0.85: base=0.80 recency=-0.00 novelty=+0.15 episodic=+0.00 pressure=+0.00 archetype=+0.030]

Persistence

The psyche is persisted in two ways:

• Session persistence: Agent::persist_session() serializes the psyche to the durable store (redb). On Agent::resume(), the persisted psyche is restored, preserving evolved weights and individuation across sessions.
• Compartment file: data/compartments/psyche/psyche.toml is the initial default, read when the psyche compartment is first loaded.

Priority order on resume: persisted session > compartment manager > default.

Configuration

Changing the persona

[persona]
name = "Explorer"
grammar_preference = "terse"
traits = ["adventurous", "bold", "direct"]
tone = ["energetic", "concise"]

Adding shadow constraints

[[shadow.veto_patterns]]
name = "no_external_network"
triggers = ["http_fetch", "curl", "wget"]
severity = 1.0
explanation = "This agent is configured for offline-only operation."

[[shadow.bias_patterns]]
name = "prefer_internal_reasoning"
triggers = ["shell_exec", "file_io"]
severity = 0.2
explanation = "Prefer KG-based reasoning over external tool calls."

Tuning archetype weights

# More exploratory
[archetypes]
healer = 0.3
sage = 0.4
guardian = 0.3
explorer = 0.9

# More cautious
[archetypes]
healer = 0.6
sage = 0.5
guardian = 0.9
explorer = 0.2

Programmatic manipulation

#![allow(unused)]
fn main() {
use akh_medu::compartment::Psyche;

let mut agent = Agent::new(engine, config)?;

let mut psyche = Psyche::default();
psyche.persona.name = "Mentor".into();
psyche.persona.grammar_preference = "formal".into();
psyche.archetypes.healer = 0.8;
agent.set_psyche(psyche);

if let Some(p) = agent.psyche() {
    println!("Dominant archetype: {}", p.dominant_archetype());
    println!("Individuation: {:.2}", p.self_integration.individuation_level);
}
}

Design Rationale

Why Jung over Asimov? Asimov's three laws are rigid boolean constraints that don't adapt. Jung's model provides a spectrum:

• The Shadow has both hard vetoes and soft biases, configurable per deployment.
• Archetypes create behavioral tendencies that evolve through experience.
• Individuation means the agent's personality matures over time.

Why is the effect subtle? The archetype bonus formula (weight - 0.5) * 0.15 produces a maximum swing of ~0.07 per tool. The psyche should nudge behavior, not override situation-specific signals. Over many cycles, the cumulative effect creates a recognizable personality.