A cross-platform memory kernel for AI coding agents. Install it on macOS, Linux, or Windows and keep project context alive across conversations.
Every time you start a new AI chat, you lose context. You waste tokens repeating instructions. Different agents don't share knowledge about your project.
Antigravity Ultra fixes this. Install 6 portable files, and every AI agent — Gemini, Claude, Codex, or any Markdown-aware tool — instantly knows your project's rules, history, and current state.
# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/ungrav/antigravity-ultra/main/install.sh | bash# Windows PowerShell
iwr https://raw.githubusercontent.com/ungrav/antigravity-ultra/main/install.ps1 -OutFile install.ps1
powershell -ExecutionPolicy Bypass -File .\install.ps1Antigravity Ultra is distributed like a small portable app, not a framework you have to assemble manually.
| Platform | Entry point | Notes |
|---|---|---|
| macOS | install.sh or portable-kernel.sh |
Bash installer, local bootstrap, portable recovery |
| Linux | install.sh or portable-kernel.sh |
Same 6-file kit and Bash bootstrap |
| Windows | install.ps1 or portable-kernel-windows.ps1 |
Native PowerShell install and bootstrap |
GitHub Releases carry the official 6-file product kit. The main branch is the source of truth; the latest release is the install channel used by install.sh and install.ps1.
Agents follow a Golden Path — they load only what the current task needs:
Normal coding → read only target files (0 extra context)
Bug fix → 10 lines from ERROR_LOG.md (~125 tokens)
Architecture → Project DNA digest (~300 tokens)
Memory work → Full memory spec (~1,940 tokens)
Every agent framework injects context per turn. Ultra's advantage isn't a smaller system prompt — it's what agents don't need to do because the project is already documented:
| Without Ultra | With Ultra |
|---|---|
| Re-explain project context every session | Agent reads current_state.md and continues |
| Manually repeat architecture decisions | Agent checks PROJECT_HISTORY.md |
| Describe past bugs and fixes | Agent reads ERROR_LOG.md |
| Re-read the entire codebase to orient | Agent reads Project DNA (50 lines) |
| Lose knowledge when switching agents | All agents share the same .agent/ state |
The Golden Path also controls additional context per task — normal coding loads zero extra files, bugs load only the error log, architecture loads only the project digest. Other frameworks load everything on every turn regardless of task type.
For project source, portable-kernel.sh project-context resolve --project-root <path> --task "<task>" maintains a local incremental filename/fingerprint index and returns a bounded read plan. Unchanged files are not rehashed, source contents and secrets are never stored, and large cross-cutting tasks are escalated to Graphify instead of repeatedly scanning every folder with the model.
Optional execution backends are discovered per machine with portable-kernel.sh backends discover. Discovery never enables delegation: the user must grant consent, and routes require current task-specific evidence. Local LM Studio and oMLX models can be benchmarked serially with explicit --allow-load; the runner preserves preloaded state, unloads only models it introduced, and persists scores and response hashes rather than prompts or response text.
The extra context an agent loads beyond the system prompt per task type:
| System | Normal task | Bug task | Architecture | Memory work |
|---|---|---|---|---|
| Antigravity Ultra | +0 | +125 | +300 | +1,940 |
| OpenCode | +661 | +661 | +661 | +661 |
| Everything Claude Code | +2,024 | +2,024 | +2,024 | +2,024 |
| Claude Code | +3,000 | +3,000 | +3,000 | +3,000 |
Over 100 turns (80 normal, 10 bugs, 5 architecture, 5 memory), that extra context totals:
| System | Extra context loaded | vs Ultra |
|---|---|---|
| Antigravity Ultra | ~12,450 | — |
| OpenCode | ~66,100 | 5× more |
| Everything Claude Code | ~202,400 | 16× more |
| Claude Code | ~300,000 | 24× more |
But the biggest saving is across sessions: with per-project memory, session 2 starts where session 1 ended. Without it, you start from zero every time.
| Feature | What it does | |
|---|---|---|
| 🧠 | Persistent Memory | Markdown Knowledge Items with structured frontmatter — agents remember decisions, bugs, and architecture |
| ⚡ | Golden Path | Task-based routing: agents load zero kernel context for 80% of tasks |
| 🤝 | Any Provider | Same rules for Gemini, Claude, Codex, or any Markdown-aware agent |
| 🕵️ | Verify, Don't Assume | Agents recheck delegated evidence, reset failed diagnoses, block PASS on open high-impact findings, and reconcile every delivery surface |
| 🧭 | Portable Backend Router | Discovers optional engines and agent CLIs per machine, asks before delegation, and routes only qualified candidates |
| 📚 | Project Ledgers | Immutable PROJECT_HISTORY.md + ERROR_LOG.md — any new agent reads them and catches up |
| 🛡️ | Safe Delete | Files go to _DEPRECATED_TRASH, never permanently deleted |
| 📐 | Zero Monoliths | Agents patch only the needed surface — no giant rewrites |
Full details → FEATURES_EN.md · FEATURES_ES.md
Everything ships as exactly 6 files:
| File | Purpose |
|---|---|
GEMINI.md |
Operational rulebook (85 lines) |
MEMORY.md |
Memory system spec (156 lines) |
GEMINI_BLUEPRINTS.md |
Recovery contract |
portable-kernel.sh |
Installer for macOS / Linux |
portable-kernel-windows.ps1 |
Installer for Windows |
minimum-kernel.bundle.tar.gz |
Runtime payload |
The installer creates everything else: .agent/, .portable/, configs, scripts. You only carry 6 files.
One command:
curl -fsSL https://raw.githubusercontent.com/ungrav/antigravity-ultra/main/install.sh | bashChoose your tier:
| Tier | What You Get |
|---|---|
🌱 minimal |
Core + portable memory + Context7 foundation |
⭐ recommended |
+ Workflows, audit tools, optional capability recommendations |
🔥 complete |
+ Telemetry and all relevant capability offers |
⚙️ custom |
Pick modules individually |
# macOS / Linux
bash portable-kernel.sh bootstrap --tier recommended
# Windows
powershell -ExecutionPolicy Bypass -File .\portable-kernel-windows.ps1 bootstrap -Tier recommendedPrerequisites: Bash 4+ or PowerShell 5+. No other dependencies.
- Not an AI model — it makes existing models smarter about your project
- Not a chatbot — it manages context and memory, not conversation UI
- Not an IDE replacement — it works inside whatever tools you already use
Open source under GPLv3. Issues and PRs are welcome.
Un kernel de memoria multiplataforma para agentes de programación con IA. Instálalo en macOS, Linux o Windows y conserva el contexto del proyecto entre conversaciones.
Cada vez que inicias un chat nuevo, pierdes el contexto. Desperdicias tokens repitiendo instrucciones. Diferentes agentes no comparten conocimiento sobre tu proyecto.
Antigravity Ultra lo resuelve. Instala 6 archivos portables y cualquier agente — Gemini, Claude, Codex o cualquier herramienta que lea Markdown — entiende al instante las reglas, historia y estado de tu proyecto.
# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/ungrav/antigravity-ultra/main/install.sh | bash# Windows PowerShell
iwr https://raw.githubusercontent.com/ungrav/antigravity-ultra/main/install.ps1 -OutFile install.ps1
powershell -ExecutionPolicy Bypass -File .\install.ps1Antigravity Ultra se distribuye como una app portable pequeña, no como un framework que tengas que ensamblar manualmente.
| Plataforma | Entrada | Notas |
|---|---|---|
| macOS | install.sh o portable-kernel.sh |
Instalador Bash, bootstrap local y recuperación portable |
| Linux | install.sh o portable-kernel.sh |
El mismo kit de 6 archivos y bootstrap Bash |
| Windows | install.ps1 o portable-kernel-windows.ps1 |
Instalación y bootstrap nativos con PowerShell |
GitHub Releases contiene el kit oficial de 6 archivos. La rama main es la fuente de verdad; el último release es el canal de instalación usado por install.sh e install.ps1.
Los agentes siguen un Golden Path — cargan solo lo que la tarea necesita:
Coding normal → solo archivos objetivo (0 contexto extra)
Bug fix → 10 líneas de ERROR_LOG.md (~125 tokens)
Arquitectura → Project DNA (~300 tokens)
Memoria → Spec completa de memoria (~1,940 tokens)
Todos los frameworks inyectan contexto por turno. La ventaja de Ultra no es un system prompt más pequeño — es lo que los agentes no necesitan hacer porque el proyecto ya está documentado:
| Sin Ultra | Con Ultra |
|---|---|
| Re-explicar contexto cada sesión | El agente lee current_state.md y continúa |
| Repetir decisiones de arquitectura | El agente consulta PROJECT_HISTORY.md |
| Describir bugs pasados | El agente lee ERROR_LOG.md |
| Re-leer todo el código para orientarse | El agente lee Project DNA (50 líneas) |
| Perder conocimiento al cambiar de agente | Todos comparten el mismo estado .agent/ |
El Golden Path también controla el contexto adicional por tarea — coding normal carga cero archivos extra, bugs cargan solo el error log, arquitectura solo el digest del proyecto. Otros frameworks cargan todo en cada turno sin importar el tipo de tarea.
Para el código del proyecto, portable-kernel.sh project-context resolve --project-root <ruta> --task "<tarea>" mantiene un índice local incremental de nombres y huellas, y devuelve un plan de lectura acotado. Los archivos sin cambios no se recalculan, nunca se guardan contenidos ni secretos, y las tareas grandes o transversales escalan a Graphify en vez de volver a recorrer todas las carpetas con el modelo.
Los backends opcionales se descubren por equipo con portable-kernel.sh backends discover. Descubrirlos nunca activa la delegación: el usuario debe consentir y cada ruta exige evidencia vigente para la tarea. Los modelos locales de LM Studio y oMLX pueden evaluarse en serie con --allow-load explícito; el runner preserva modelos ya cargados, descarga solo los que él introdujo y guarda puntajes y hashes de respuesta, no prompts ni respuestas.
El contexto extra que un agente carga más allá del system prompt por tipo de tarea:
| Sistema | Tarea normal | Bug | Arquitectura | Memoria |
|---|---|---|---|---|
| Antigravity Ultra | +0 | +125 | +300 | +1,940 |
| OpenCode | +661 | +661 | +661 | +661 |
| Everything Claude Code | +2,024 | +2,024 | +2,024 | +2,024 |
| Claude Code | +3,000 | +3,000 | +3,000 | +3,000 |
En 100 turnos (80 normales, 10 bugs, 5 arquitectura, 5 memoria), ese contexto extra suma:
| Sistema | Contexto extra cargado | vs Ultra |
|---|---|---|
| Antigravity Ultra | ~12,450 | — |
| OpenCode | ~66,100 | 5× más |
| Everything Claude Code | ~202,400 | 16× más |
| Claude Code | ~300,000 | 24× más |
Pero el mayor ahorro es entre sesiones: con memoria por proyecto, la sesión 2 empieza donde terminó la sesión 1. Sin ella, empiezas de cero cada vez.
| Característica | Qué hace | |
|---|---|---|
| 🧠 | Memoria Persistente | Knowledge Items en Markdown con frontmatter estructurado — los agentes recuerdan decisiones, bugs y arquitectura |
| ⚡ | Golden Path | Enrutamiento por tarea: cero contexto del kernel para el 80% de las tareas |
| 🤝 | Cualquier Proveedor | Mismas reglas para Gemini, Claude, Codex o cualquier agente Markdown |
| 🕵️ | Verificar, No Asumir | Los agentes corroboran evidencia delegada, reinician diagnósticos fallidos, bloquean PASS por findings graves y reconcilian toda la entrega |
| 🧭 | Router Portable de Backends | Descubre motores y CLIs de agentes disponibles en cada equipo, pide permiso antes de delegar y sólo enruta candidatos calificados |
| 📚 | Historial Permanente | PROJECT_HISTORY.md + ERROR_LOG.md inmutables — cualquier agente nuevo se pone al día |
| 🛡️ | Borrado Seguro | Los archivos van a _DEPRECATED_TRASH, nunca se eliminan permanentemente |
| 📐 | Cero Monolitos | Los agentes parchean solo la superficie necesaria |
Detalles completos → FEATURES_ES.md
Todo se distribuye en exactamente 6 archivos:
| Archivo | Propósito |
|---|---|
GEMINI.md |
Reglamento operativo (85 líneas) |
MEMORY.md |
Especificación de memoria (156 líneas) |
GEMINI_BLUEPRINTS.md |
Contrato de recuperación |
portable-kernel.sh |
Instalador para macOS / Linux |
portable-kernel-windows.ps1 |
Instalador para Windows |
minimum-kernel.bundle.tar.gz |
Payload runtime |
El instalador crea todo lo demás: .agent/, .portable/, configs, scripts. Solo cargas 6 archivos.
Un solo comando:
curl -fsSL https://raw.githubusercontent.com/ungrav/antigravity-ultra/main/install.sh | bashElige tu perfil:
| Perfil | Qué incluye |
|---|---|
🌱 minimal |
Motor + memoria portable + Context7 base |
⭐ recommended |
+ Workflows, auditoría y recomendaciones de capacidades opcionales |
🔥 complete |
+ Telemetría y todas las capacidades relevantes ofrecidas |
⚙️ custom |
Elige módulos individualmente |
# macOS / Linux
bash portable-kernel.sh bootstrap --tier recommended
# Windows
powershell -ExecutionPolicy Bypass -File .\portable-kernel-windows.ps1 bootstrap -Tier recommendedRequisitos: Bash 4+ o PowerShell 5+. Sin otras dependencias.
- No es un modelo de IA — hace que los modelos existentes sean más inteligentes con tu proyecto
- No es un chatbot — gestiona contexto y memoria, no interfaces de conversación
- No reemplaza tu IDE — funciona dentro de las herramientas que ya usas
Open source bajo GPLv3. Issues y PRs son bienvenidos.
GNU GPLv3 — Ver LICENSE.