Builder of AI agent infrastructure. I design protocols and systems that let agents remember, orient, and grow — without databases, embeddings, or cloud services. Just files.
Currently building Soma — an AI coding agent with self-growing memory — and publishing the protocol specs behind it as open standards.
Gravicity is the ecosystem. Soma is the first product.
🌿 SomaAI coding agent with memory Built on Pi. Identity discovery, session breathing, filesystem memory, protocol-driven behavior. The agent that grows around you.
|
Open specs for agent architecture Five protocol specs born from building Soma — published as standalone standards anyone can implement: |
| Protocol | What It Does |
|---|---|
| AMP | Agent Memory Protocol — filesystem-based persistent memory with muscles, preloads, heat tracking |
| ATLAS | Architecture Truth Layered Across Stacks — living system maps that stay accurate |
| Breath Cycle | Session lifecycle: inhale → process → exhale → rest. Context depletion as design, not bug |
| Three-Layer Model | Extensions (code) + Skills (knowledge) + Rituals (workflows) — separation of agent capabilities |
| Identity System | Contextual identity discovery — agents that know who they are based on where they are |
All specs: CC BY 4.0 · Reference implementation: Soma (MIT)
Most frameworks treat agent memory as a retrieval problem — vector databases, embeddings, RAG.
I think it's simpler than that. The agent reads and writes files. Like a human with a notebook.
No infrastructure. No embeddings model. No cloud service. Just Markdown on disk.
The context window isn't a problem to solve. It's a breath.
It fills, it empties, it fills again. An agent that knows it will forget is more capable than one that pretends it won't.
- openclaw-mods — Community patches for OpenClaw: per-agent compaction, context management
- acpx — Headless CLI client for stateful Agent Client Protocol sessions
- polymarket-tui — Copy, scalping & sniper TUI for Polymarket
soma.gravicity.ai · gravicity.ai
Systems that build systems.