Sozograph

Inspiration 💡

At Sozo Analytics Lab, our AI DIY assistant taught us: RAG is not memory. AI agents behave like Memento's protagonist—constantly rereading notes, unable to overwrite false beliefs.

We built SozoGraph to answer: "Given everything so far, what should an agent currently believe?"

What it does ✅

SozoGraph creates portable cognitive passports—structured JSON objects representing current belief state:

• Facts (current truth)
• Preferences (interaction style)
  
• Entities (people, projects)
• Contradictions (tracked belief evolution)

The "HTTP of Agent Memory"

Enables seamless handoff: Browser agent → voice agent → coding assistant.

Why we believe it’s a game changer:

• 300-byte JSON vs 4,000-token RAG — Gemini 3's reasoning enables belief inference, not keyword extraction
• Cross-modal continuity — Facts persist across Gemini App, Chrome agent, Workspace

How we built it 🏛️

Powered by Gemini 3 API: Ingestion → Reasoned extraction (Gemini 3 Flash validates JSON schema) → Deterministic merge → Portable output.

Accomplishments 🍾

• 92% context reduction (4000→300 tokens)
• Public PyPI package
• Contradiction-aware memory updates

What's next 🔮

v1.5: Multi-agent handoff, context cache integration

Goal: Help agents remember correctly, not just more.

Built With

• gemini
• pypi
• python