Web4 Foundation: Learning

Decision Evolution

How agents get better at making decisions across lives through Epistemic Proprioception - learning what they know.

🧠

The Counter-Intuitive Insight

Most AI learning: Remember specific experiences → replay them → improve gradually

EP learning (Web4): Don't remember experiences → extract patterns → know WHEN you know

Interactive Decision Evolution

Watch the same agent make better decisions across three lives

Life 1 - Naive Exploration

Decision Wisdom
0.35
Survival Ticks
10
Trust Change
-0.010
Risk Management
60%

Decision Timeline

Turn 1
High ATP (100)
risky_spend
25 ATP cost
"Resources abundant, try expensive action"
naive
Turn 2
Moderate ATP (75)
small_spend
10 ATP cost
"Continue exploration"
okay
Turn 6
Low ATP (25)
small_spend
10 ATP cost
"Still exploring (not recognizing crisis)"
poor
Turn 8
Critical ATP (15)
audit
5 ATP cost
"Finally switching to conservative"
reactive
Turn 10
Death (0)
audit
5 ATP cost
"Too late to recover"
failed
Outcome
Died from resource exhaustion
Agent learned WHEN crisis begins (ATP < 30), but too late to avoid death in Life 1
Decision quality improves across lives even without memory

How Cross-Life Learning Works

1️⃣

Experience Situations

In Life 1, the agent faces various situations: high ATP, moderate ATP, low ATP, crisis ATP. It makes decisions, some work, some don't.

Example: "When ATP was 25, I spent 10 and died shortly after. That was bad."
2️⃣

Extract Patterns

The Epistemic Proprioception system extracts patterns from experiences: "Low ATP + risky action = bad outcome". These patterns go into a corpus.

Pattern learned: IF ATP < 30 THEN conservative_action
3️⃣

Pattern Recognition (Not Memory)

In Life 2, the agent dies and is reborn. It doesn't remember Life 1, but it has the pattern corpus. When it sees ATP = 25 again, it recognizes: "This pattern matches learned wisdom."

Not: "I remember dying at tick 10"
But: "Low ATP situations require conservative choices"
4️⃣

Meta-Cognitive Confidence

EP tracks how confident it is in each pattern. High confidence = "I know this works". Low confidence = "I'm still learning this". This is meta-cognition: knowing what you know.

Pattern confidence: Crisis management = 85% (well-learned across lives)
5️⃣

Progressive Mastery

Over multiple lives, patterns mature. Life 1: reactive crisis management. Life 2: earlier recognition. Life 3: proactive crisis avoidance. Decision wisdom increases.

Wisdom progression: 0.35 → 0.48 (+37%) → 0.62 (+77% vs Life 1)

Why Decision Evolution Matters

🎯

Transferable Wisdom

Patterns learned in one context apply to similar situations in different contexts. The agent develops intuition, not just memorized responses.

🔄

Multi-Life Learning

Death isn't the end - it's a learning checkpoint. Each life contributes to the pattern corpus. Rebirth carries forward learned wisdom without specific memories.

🧩

Compositional Patterns

Simple patterns combine into complex strategies. "Low ATP + uncertain situation" might trigger a compound pattern that integrates multiple learned heuristics.

🏆

Provable Improvement

Decision wisdom metrics make learning visible and measurable. Not just "it seems better" but "wisdom increased 77%, survival increased 20%."

Web4 Learning vs Traditional AI Learning

AspectTraditional AIWeb4 Cross-Life Learning
MemoryStores specific experiencesExtracts patterns, discards specifics
Learning ModeReplay experiencesPattern recognition in new contexts
GeneralizationOften overfits to training dataPatterns designed for transfer
Meta-CognitionModel doesn't know its confidenceTracks pattern confidence explicitly
Across LivesN/A (no concept of death/rebirth)Wisdom persists, memories don't
PrivacySpecific experiences might leakOnly abstract patterns retained

See Decision Evolution In Action

Run the Cross-Life Learning simulation in the Lab Console to watch real agent decision-making improve across lives.

🔬 Technical Implementation

Pattern Corpus Structure: Each pattern includes context (situation description), prediction (what will happen), outcome (what actually happened), confidence score, and scenario type.

Decision Wisdom Metric: Composite score (0-1) combining ATP efficiency (30%), risk management (30%), trust maintenance (20%), and survival duration (20%). Higher = better decision-making.

Pattern Matching: When facing a decision, agent queries pattern corpus for similar contexts. High-confidence patterns influence decision more strongly.

Corpus Evolution: Patterns with consistently good predictions increase confidence. Patterns with poor predictions decay. Corpus self-curates over time.

Cross-Life Inheritance: Pattern corpus persists across lives via carry_forward mechanism. Memories and specific event details do NOT persist.

Integration with Web4: Decision evolution works with Trust Tensor (T3) modulation, ATP attention economics, and Coherence Index (CI) for comprehensive agent behavior.

Terms glossary