EvAGI

Documentation · not source

The experimental system behind EvSphere's continual-learning research

Website · EvSphere

---

What this repository is. Documentation for EvAGI — what it does, how it is organised, and what state it is in.

What it is not. This is not the source repository. No implementation code, configuration, prompts or infrastructure lives here, and none will.

---

❓ Why it exists

Zero measured forgetting in both published runs — and the reason matters. Each task owns disjoint parameters, so there is no shared weight for one task to overwrite. Isolation removes interference by construction. That is a real result with a narrow claim, not a solution to catastrophic forgetting.

---

🧩 What it does

• Routes each input to the expert set that should handle it
• Keeps experts and their parameter pools disjoint, so no weights are shared
• Predicts how much capacity a task needs rather than activating everything
• Learns facts live and re-checks them in a fresh session
• Reports measured accuracy, forgetting and capacity alongside every run

---

🔄 How it fits together

         Input
            ↓
         Router                 ← decides from a live support set
            ↓
  ┌─────────┼─────────┐
  ▼         ▼         ▼
Expert A Expert B Expert C     ← disjoint, no shared parameters
  │         │         │
  ▼         ▼         ▼
Pool A    Pool B    Pool C     ← each owns its own parameters
  │         │         │
  └─────────┼─────────┘
            ↓
          Merge
            ↓
         Result

---

📊 Status

Maturity	Experimental — see STATUS.md
Version	research prototype
Public page	https://evsphere.vercel.app/research/evagi

---

⚠️ Known limits

• Catastrophic forgetting is NOT solved by this work
• Zero measured forgetting arises because each task gets disjoint parameters — interference is removed by construction, not by improving retention
• Both published runs are single runs — no seeds, repeats or error bars
• Neither task set is a standard continual-learning benchmark
• Nothing is released publicly, so results cannot be independently reproduced
• The efficiency comparison is against full fine-tuning, the weakest available baseline

Full detail, including what is not claimed, is in docs/STATUS.md.

---

📚 Documentation

Document	Covers
ARCHITECTURE.md	Conceptual structure and subsystem responsibilities
docs/HOW-IT-WORKS.md	Stage by stage, and where the boundaries are
docs/STATUS.md	Maturity, requirements and what is not claimed
docs/FAQ.md	Common questions and honest answers
NOTICE.md	What is deliberately absent, and licensing

---

🔗 EvSphere

🌐 Website · GitHub · npm · LinkedIn · X

Machine-readable: llms.txt · llms-full.txt

---

⚖️ Disclosure

Internal implementation detail is not published here. No credential, endpoint, key, dataset or client material is included — see NOTICE.md.

If you find any of those here, it is an error. Report it privately through the enquiry page rather than opening a public issue.