Skip to content

Home

Jump to bottom
Yash Kanadhia edited this page Jun 9, 2026 · 4 revisions

Zeref OS Wiki

Zeref OS

Local-first context and memory engine for AI-assisted work. Harness-agnostic · Model-agnostic · Privacy-first · Free

Named after Zeref Dragneel from Fairy Tail — the immortal scholar who carried ancient knowledge across forms and ages. Zeref OS is built in that lineage: long-horizon memory, faithful to the user's accumulated decisions, portable across every AI harness.

Current version: v2.6.1 (June 8, 2026) — CHANGELOG · docs/RELEASE_LOG.md · Notion Command Center

Rubric: 9.88/10 (8 dims, every score cites artifact) — tests/zeref-rubric-v2.6.md

Quick links

  • 📦 Installation — per-harness setup, verification, uninstall
  • 🏗️ Architecture — 6 agents · 14 skills · 8 commands · 6 team packs · 4 Auto-Activation Gates · Model-Tier Routing
  • 🧠 Memory Model — flat layout, boundary-first reads, contradiction handling, PATTERNS.jsonl event schema (L5+L15)
  • 🔒 Privacy Model — PRIVACY/REDACT/SHARING_POLICY, modes, connectors, R6 Zero Context Loss
  • 👥 Team Packs — solo / build / research / red / audit / ship
  • 🔍 Pattern Detection — Two-Strikes Rule, pattern-observer, skill drafting
  • 📜 Decision Log — D1–D11 + v2.6 + v2.6.1 arbitrations
  • 🤖 Model Debates — what Claude / GPT / Gemini / open-source each need + v2.6 model-resolver
  • 🕰️ Versioning History — Skills Fleet → Agent OS → Zeref OS → v2.5 audit → v2.6 4-gate → v2.6.1 hardening
  • ❓ FAQ — common questions
  • 📖 Glossary — boundary file, evidence grade, Two-Strikes, R6, 4-gate chain, model-resolver, etc.
  • 🌱 Inspirations — engineering lineage and influences

What Zeref OS is (one paragraph, v2.6.1)

Per-project flat memory/ wiki in plain markdown, append-only PATTERNS.jsonl event log with schema validator, point-in-time snapshots, contradiction safety with human arbitration, three privacy modes (exact / abstract / local-only), six on-demand team packs, cross-harness handoff format with caveman-grammar compression, and a 4-gate auto-activation chain (budget-governor → skill-router → fleet-activator → prompt-context-engine) that classifies every major task on cost weight, picks the smallest useful skill stack, probes extended-tool reachability, and restructures unstructured prompts before any token spend. AI sessions become cumulative across every harness you use.

v2.6 four-gate chain (every major task)

[budget-governor]      classify weight (CRITICAL/HIGH/MEDIUM/LOW) + match model tier
       ↓
[skill-router]         pick smallest stack (1 lead + 2-3 support + 1 QA, max 5 skills)
       ↓
[fleet-activator]      live-probe ECC / claude-obsidian / Graphify / browser-harness / notebooklm / gstack
       ↓
[prompt-context-engine] classify STRUCTURED / SEMI-STRUCTURED / UNSTRUCTURED; restructure if needed; R6 zero context loss
       ↓
                       execute (declared stack, declared brief, declared tier)
       ↓
[caveman-handoff]      compress cross-model handoff (40-60% reduction; NFKC + R6 diff)

Each gate declares its result inline. User can override before token spend. Per AGENTS.md ## Auto-Activation Gates.

Where to start

If you want to... Read
Install in 5 minutes Installation
Understand the system Architecture → Memory Model
Lock down privacy first Privacy Model
See how teams work Team Packs
Trace design decisions Decision Log
Trace the iteration history Versioning History
Understand who we built on Inspirations
Run the audit yourself scripts/zeref-validate.py — Skills 14/14, PATTERNS lint 0

README · AGENTS.md · CHANGELOG · GITHUB_OS.md · docs/adr/ · _shared/model-resolver.md

Zeref OS

Getting Started

Architecture

Reference

Clone this wiki locally