Turing Pyramid

Prioritized action selection for AI agents. 10 configurable needs with time-based decay, tension scoring, and weighted action selection. Hook it into your heartbeat — get concrete "do this next" suggestions instead of idle loops.

The Problem

Without structure, agents either:

Idle — wait for prompts, do nothing between interactions
Spam — check the same thing every cycle, repeat low-value actions
Drift — pick random tasks with no prioritization

Default existence cycles nudge agents with "do what feels right" — but without state tracking, the agent has no memory of what it already did, what's been neglected, or what matters most right now.

Turing Pyramid replaces that with a stateful feedback loop: needs decay over time → tension builds → highest-tension need gets an action → satisfaction resets → cycle continues. It works as a drop-in replacement for the native OpenClaw existence cycle, with actual prioritization and action variety built in.

What Changes For Your Agent

Before (typical heartbeat):

Heartbeat → "anything to do?" → nothing obvious → HEARTBEAT_OK
Heartbeat → "anything to do?" → check inbox again → HEARTBEAT_OK
Heartbeat → "anything to do?" → HEARTBEAT_OK

After (with Turing Pyramid):

Heartbeat → coherence tension=16, closure=14, connection=10
  → ACTION: sync daily logs to MEMORY.md (coherence, impact 1.8)
  → ACTION: complete one pending TODO (closure, impact 1.7)
  → NOTICED: connection — deferred

Heartbeat → connection tension=12, expression=8, understanding=6
  → ACTION: reply to pending mentions (connection, impact 1.8)
  → ACTION: write journal reflection (expression, impact 1.8)
  → NOTICED: understanding — deferred

The agent rotates through different types of work based on what's been neglected longest.

How It Works

10 needs, each with configurable importance (priority weight) and decay rate (how fast satisfaction drops):

Need	Importance	Decay	What it tracks
security	10	168h	Backups, vault integrity, system health
integrity	9	72h	Behavior aligned with stated values
coherence	8	24h	Memory organization, no contradictions
closure	7	12h	Open tasks and threads getting resolved
autonomy	6	36h	Self-initiated decisions and projects
connection	5	8h	Social interaction, community participation
competence	4	36h	Successful task completion, skill growth
understanding	3	12h	Learning, research, curiosity
recognition	2	48h	Sharing work, getting feedback
expression	1	8h	Writing, creating, articulating thoughts

Each cycle:

Satisfaction decays based on elapsed time (0.0–3.0 range)
Tension = importance × deprivation — higher = more urgent
Top 3 needs by tension get action slots
Probability roll decides action vs. notice (higher tension = higher chance)
Impact matrix selects action size (crisis → big actions, stable → small maintenance)
Weighted random picks specific action from the selected impact range
Cross-need effects propagate (e.g., completing a task boosts both closure and competence)

Protection mechanisms:

Starvation guard — any need stuck at floor for 48h+ gets a forced action slot
Action staleness — recently-picked actions get weight penalty to prevent repetition
Follow-ups — temporal markers to check results of past actions ("posted on Moltbook → check replies in 4h")
Day/night decay — configurable multiplier for different time periods
Floor/ceiling — satisfaction clamped to 0.5–3.0, prevents runaway states

Quick Start

# Initialize state file
./scripts/init.sh

# Add to HEARTBEAT.md:
<skill-dir>/scripts/run-cycle.sh

# After completing a suggested action:
./scripts/mark-satisfied.sh <need> [impact]

# With follow-up (check back later):
./scripts/mark-satisfied.sh connection 1.5 --reason "posted update" --followup "check replies" --in 4h

# Manual follow-up (e.g., from steward):
./scripts/create-followup.sh --what "review PR CI" --in 2h --need competence --source steward

Requires: bash, jq, bc, grep, find + WORKSPACE env var set.

Customization

Everything is in assets/needs-config.json:

Decay rates — how fast each need builds tension
Action lists — what gets suggested per need (add your own)
Weights — probability of each action being selected
Importance — which needs win when multiple compete
Disable a need — set importance: 0

Guided onboarding conversation template included in SKILL.md.

See references/TUNING.md for detailed tuning guide.

Architecture: Suggestion Engine, Not Executor

The skill outputs text suggestions. It does not execute actions, make network requests, or access credentials.

Turing Pyramid (local-only)      Your Agent (has capabilities)
───────────────────────────      ─────────────────────────────
reads JSON config + state    →   receives "★ do X" text
scans workspace files        →   decides: execute? skip? ask human?
outputs suggestion text      →   uses its own tools and permissions

Reads: workspace files (MEMORY.md, SOUL.md, etc.) via grep/find for pattern detection. Writes: assets/needs-state.json only (timestamps and satisfaction levels). Never accesses: credentials, APIs, network, paths outside workspace.

Token Usage

Heartbeat interval	Cycles/day	Est. tokens/day	Est. tokens/month
30 min	48	48k–120k	1.4M–3.6M
1 hour	24	24k–60k	720k–1.8M
2 hours	12	12k–30k	360k–900k

Stable agents (most needs satisfied) use fewer tokens. First few days are higher as the system stabilizes.

Files

turing-pyramid/
├── SKILL.md              # Full documentation
├── DESCRIPTION.md        # This file
├── assets/
│   ├── needs-config.json # ★ Needs, decay rates, actions — tune this
│   ├── needs-state.json  # Runtime state (auto-managed)
│   ├── followups.jsonl   # Follow-up markers (auto-managed)
│   └── cross-need-impact.json  # Inter-need effects
├── scripts/
│   ├── run-cycle.sh      # Main heartbeat entry point
│   ├── mark-satisfied.sh # Update state after action (supports --followup)
│   ├── create-followup.sh # Create temporal check-back markers
│   ├── resolve-followup.sh # Close follow-ups (single or bulk)
│   ├── show-status.sh    # Debug current tensions
│   ├── init.sh           # First-time state setup
│   └── scan_*.sh         # 10 workspace scanners
├── tests/                # 50+ test cases (unit + integration)
└── references/
    ├── TUNING.md         # Customization guide
    └── architecture.md   # Technical deep-dive

7.3 KiB Raw Blame History Unescape Escape