Identiti / Aegis

Your memory lives here. Chat with an AI that knows who you are — and builds your knowledge graph as you talk.

Docs: PROJECT_OVERVIEW.md (architecture) · benchmarks/EVAL_REPORT.md (eval numbers) · benchmarks/README.md (reproduce)

Architecture

flowchart TD
    User("User")

    User -->|Google OAuth| Supabase("Supabase Auth")
    User --> Frontend

    subgraph Frontend ["React + Vite"]
        direction LR
        Login("Login") --> Onboarding("Onboarding") --> Chat("Chat") --> Graph("Graph")
    end

    Frontend -->|JWT| Flask

    subgraph Flask ["Flask API"]
        direction TB
        ChatAPI("/chat") --> LLM
        ChatAPI -->|extract entities| LLM
        OnboardChat("/api/onboard-chat") --> LLM
        Import("/api/import") --> LLM
        NodeAPI("/api/graph/node") --> LLM
        LLM("Groq\nllama-3.3-70b")
    end

    ExternalAI("ChatGPT / Claude") -->|memory export| Import

    Flask -->|read & write| Neo4j[("Neo4j\nKnowledge Graph")]
    Flask -->|conversation history| Redis[("Redis")]
    Supabase -->|verify JWT| Flask

Loading

Stack

Layer	Tech
Frontend	React + Vite, canvas graph renderer
Backend	Flask (Python)
Auth	Supabase (Google OAuth)
LLM	Groq — llama-3.3-70b via litellm
Graph DB	Neo4j
Cache	Redis (conversation history)

Where AI is called

Endpoint	Purpose
/chat	Generate replies using memory context
/chat (post-reply)	Extract entities from conversation → graph nodes
/api/onboard-chat	Onboarding conversation → profile JSON
/api/import	Parse memory export → structured profile
/api/graph/node	Normalize user-typed labels (spelling correction)
/api/wallet	(reads graph, no LLM)

Run it

python3 app.py
# open http://localhost:3000

Env vars

GROQ_API_KEY
LLM_MODEL          # default: groq/llama-3.3-70b-versatile
NEO4J_URI
NEO4J_USER
NEO4J_PASSWORD
NEO4J_DATABASE
SUPABASE_URL
SUPABASE_JWT_SECRET
REDIS_URL           # optional

Benchmarks

LongMemEval-S harness in benchmarks/. Dev slice (n=10): Aegis QA accuracy 80% vs flat BM25 RAG 70% at recall@10 parity — see EVAL_REPORT.

pip install -r requirements-benchmark.txt
python3 benchmarks/download_data.py
python3 benchmarks/run_retrieval.py --backend flat_rag --limit 5   # smoke test
python3 benchmarks/verify_reference.py --run-id dev                # vs committed reference