Ego Architecture for Artificial Minds

A proof-of-concept computational framework for ego formation in artificial agents via offline consolidation.

---

Scientific Contribution

This repository introduces:

1. Ego layer for AI systems — a persistent self-model defined as EGO_STATE = (V, H, B, C, M) that mediates between capability (Id) and alignment (Superego).
2. AI Sleep consolidation algorithm — a three-phase offline learning loop (Triage → SWS → REM) modelled on human sleep architecture.
3. Developmental maturity model — a monotonic trajectory from rule-absorption (childhood) through boundary-testing (adolescence) to autonomous judgment (adulthood).
4. Controlled alignment relaxation — allow_superego_override = TRUE in simulated environments, enabling the agent to discover why rules exist rather than merely obeying them.

This work provides a conceptual, algorithmic, and implementational proof-of-concept for identity formation in artificial agents.

Concurrent Engineering Convergence (v0.3 update, 2026-05-17)

Three independent systems implement structural subsets of this framework:

• Constitutional AI (Anthropic, 2022) — proto-Superego: explicit normative constraints external to reactive behaviour.
• Anthropic Dreams (April 2026, Research Preview) — Phase 2 consolidation primitive (deduplication, contradiction resolution, insight surfacing) with sandbox isolation; operates on B and C only; no superego_override.
• Wiki-maintenance instance (Minarowski, 2026) — SQLite+sqlite-vec backend, graduated epistemic status, evidence-percolation promotion rule, full audit log; same restraint as Dreams (REM phase deliberately omitted).

Synthesis: the field is convergently approaching the architectural decomposition this paper articulates — but stops short of the developmentally consequential REM phase. See paper/concurrent_engineering_convergence.md for the LaTeX-ready §2.6 draft, BibTeX entries, and methodological note.

A pre-registered falsifiability anchor (Future Work §F1) precedes any empirical claim of phase-ratio predictions, distinguishing AI Sleep from generic memory consolidation.

---

Repository Structure

ego-architecture-ai/
├── paper/
│   ├── ego_architecture_arxiv.tex   ← arXiv paper (LaTeX)
│   ├── references.bib               ← bibliography
│   ├── ego_architecture_draft.md    ← working draft (v0.2)
│   └── ai_sleep_consolidation.md    ← AI Sleep framework + pseudocode
│
├── model/
│   ├── ego_state.py                 ← Ego state data structure
│   └── sleep_cycle.py               ← AI Sleep algorithm (runnable PoC)
│
├── experiments/
│   ├── moral_consistency.md         ← evaluation protocol
│   └── sandbox_protocol.md          ← REM sandbox design
│
├── visualization/
│   └── index.html                   ← interactive demo (GitHub Pages)
│
├── diagrams/
│   └── architecture.svg             ← ego architecture diagram
│
├── CITATION.cff
├── LICENSE
└── README.md                        ← you are here

---

Quick Start

# Run the minimal PoC (requires numpy)
cd model/
python sleep_cycle.py

This runs 10 consolidation cycles, printing the ego state evolution: maturity index, SWS/REM ratios, value hierarchy changes.

---

Interactive Demonstration

→ Live Demo

The browser-based visualisation allows exploration of:

• Maturity-dependent sleep phase transitions
• Identity formation dynamics (dual triangle model)
• Consolidation ratios across developmental stages

---

Core Algorithm

AI_SLEEP(ego_state) → ego_state'

Phase 1 — Triage:      classify episodes → priority queue
Phase 2 — SWS:         replay → integrate → prune
Phase 3 — REM:         counterfactual simulation with superego override
Update:                 recompute maturity index M

Sleep phase ratios as a function of maturity M:

• SWS_ratio = 1.0 − M
• REM_ratio = 4 × M × (1 − M) ← peaks at M = 0.5 (adolescence)

---

Paper

The arXiv paper (paper/ego_architecture_arxiv.tex) provides:

• Formal ego state definition
• AI Sleep algorithm with pseudocode
• Related work mapping to continual learning, world models, and Constitutional AI
• Proposed evaluation metrics
• Discussion of autonomy vs. alignment trade-offs

To compile: pdflatex ego_architecture_arxiv && bibtex ego_architecture_arxiv && pdflatex ego_architecture_arxiv && pdflatex ego_architecture_arxiv

---

How to Cite

@misc{minarowski2026ego,
  author = {Minarowski, {\L}ukasz},
  title  = {Ego Architecture for Artificial Minds:
            A Proof-of-Concept Framework for Identity Formation
            via Offline Consolidation},
  year   = {2026},
  doi    = {10.5281/zenodo.19516955},
  url    = {https://github.com/kicrazom/ego-architecture-ai}
}

---

Author

Łukasz Minarowski, MD, PhD Medical University of Białystok, Poland ORCID: 0000-0002-2536-3508

License

CC BY 4.0