Summary
The Turing Pyramid (GitHub) is a published OpenClaw skill implementing a full psychological-needs engine for autonomous agents. It's worth understanding deeply before deciding whether any of it belongs in Ei.
Background
We originally referenced this in #34 thinking it was just "surface topic tension to heartbeat." It's significantly more than that — and also probably not a drop-in fit for Ei as-is. This issue captures what we learned and leaves a door open for future work.
What It Actually Is
"Prioritized action selection engine using 10 configurable needs with time-decay and tension scoring. Each cycle calculates which needs are most deprived and selects concrete actions ranked by urgency."
The 10 Needs (universal across agents, not per-persona)
| Need |
Importance |
Decays every |
Character |
| Security |
10 |
7 days |
System stability, data integrity |
| Integrity |
9 |
3 days |
Alignment between behavior and stated principles |
| Coherence |
8 |
1 day |
Memory organization, internal consistency |
| Closure |
7 |
12 hours |
Completion of started tasks and threads |
| Autonomy |
6 |
36 hours |
Self-initiated action, decisions from own values |
| Connection |
5 |
8 hours |
Social bonds, community participation |
| Competence |
4 |
36 hours |
Mastery, successful task completion |
| Understanding |
4 |
8 hours |
Intellectual curiosity and learning |
| Recognition |
2 |
2 days |
Feedback, acknowledgment of contribution |
| Expression |
1 |
8 hours |
Externalization of thoughts for clarity |
Tension Formula ("Turing-exp")
tension = dep² + importance × max(0, dep - crisis_threshold)²
dep = deprivation = 3.0 - satisfaction (satisfaction ranges 0.5–3.0)- Below crisis (dep < 1.0): all needs compete equally regardless of importance
- Above crisis (dep > 1.0): importance amplifies the crisis signal — security gets 3.4× priority over expression at max deprivation
Time-Based Decay (not mention-frequency)
Satisfaction drops on a clock, not based on conversation. Connection (8h decay) depletes whether discussed or not. Day/night multiplier slows decay 50% at night.
Cross-Need Cascades (11 directed edges)
Satisfying one need partially satisfies others. Starving one drags others down.
| Source → Target |
On Action |
On Deprivation |
| expression → recognition |
+0.25 |
— |
| connection → expression |
+0.20 |
−0.15 |
| competence → recognition |
+0.30 |
−0.20 |
| autonomy → integrity |
+0.20 |
−0.25 |
| closure → coherence |
+0.20 |
— |
Other Components
- Action selection: probability table by satisfaction level (100% at crisis → 0% at perfect), weighted random from impact bands
- Execution gate: structural enforcement — the system verifies actions actually happened (file created, command run) before marking satisfied
- 3-layer spontaneity system: organic behavior variation, prevents mechanical cycling
- MINDSTATE continuity layer: cron-based daemon maintaining state across sessions, boot reconciliation
Why It's Not a Direct Fit for Ei
The Turing Pyramid is designed for autonomous agents running on cron jobs — the agent wakes up, runs a cycle, takes an action, goes back to sleep. Ei personas are embedded in a conversation system — they respond to messages and fire heartbeats, but they don't independently execute file operations or run shell commands.
The execution gate and MINDSTATE continuity are the most architecturally foreign pieces. Ei already has its own continuity (StateManager + ceremony) and its own action model (LLM responses + tool calls).
What Might Actually Fit
The needs themselves, the decay model, and the tension formula are interesting. They represent a principled answer to "what should a persona want right now?" that goes beyond what Ei's current exposure gap provides.
Specific pieces worth exploring:
- Universal needs as a second layer beneath PersonaTopics: topics = what the persona cares about; needs = what they're hungry for right now. These are orthogonal.
- Time-based satisfaction decay vs. Ei's mention-frequency exposure model: could Ei's heartbeat be richer if "connection" decays on a clock instead of requiring explicit message counting?
- Cross-need cascades as a way to model how persona state affects character: a persona starved of "integrity" behaves differently than one at equilibrium.
Open Questions
- Do the 10 needs map cleanly onto Ei's Persona model, or would they need to be per-persona and configurable?
- Is the execution gate concept applicable to tool-calling personas? ("Did the persona actually use the Spotify tool or just say it would?")
- Would the MINDSTATE continuity pattern add anything Ei's ceremony doesn't already provide?
- What would it cost to run the tension formula on each heartbeat, and is the output meaningfully different from the current exposure gap?
References
Summary
The Turing Pyramid (GitHub) is a published OpenClaw skill implementing a full psychological-needs engine for autonomous agents. It's worth understanding deeply before deciding whether any of it belongs in Ei.
Background
We originally referenced this in #34 thinking it was just "surface topic tension to heartbeat." It's significantly more than that — and also probably not a drop-in fit for Ei as-is. This issue captures what we learned and leaves a door open for future work.
What It Actually Is
The 10 Needs (universal across agents, not per-persona)
Tension Formula ("Turing-exp")
dep= deprivation =3.0 - satisfaction(satisfaction ranges 0.5–3.0)Time-Based Decay (not mention-frequency)
Satisfaction drops on a clock, not based on conversation. Connection (8h decay) depletes whether discussed or not. Day/night multiplier slows decay 50% at night.
Cross-Need Cascades (11 directed edges)
Satisfying one need partially satisfies others. Starving one drags others down.
Other Components
Why It's Not a Direct Fit for Ei
The Turing Pyramid is designed for autonomous agents running on cron jobs — the agent wakes up, runs a cycle, takes an action, goes back to sleep. Ei personas are embedded in a conversation system — they respond to messages and fire heartbeats, but they don't independently execute file operations or run shell commands.
The execution gate and MINDSTATE continuity are the most architecturally foreign pieces. Ei already has its own continuity (StateManager + ceremony) and its own action model (LLM responses + tool calls).
What Might Actually Fit
The needs themselves, the decay model, and the tension formula are interesting. They represent a principled answer to "what should a persona want right now?" that goes beyond what Ei's current exposure gap provides.
Specific pieces worth exploring:
Open Questions
References