Getting Started

Architecture

Seren is built around a handwritten ReAct loop — no LangChain. The agent reasons, acts against your connected tools, observes results, and iterates until it has a root cause or exhausts its tool budget.

ReAct Loop
Reason → Act → Observe, repeated until the agent reaches sufficient confidence. Every step is logged and surfaced in the investigation timeline.

Stack

stack
Frontend:   Next.js 14 + Zustand
Backend:    FastAPI + ARQ/Redis job queue
Agent:      Handwritten ReAct loop + Groq (free tier)
Memory:     Redis + Qdrant + Neo4j
Auth:       Supabase Auth
Database:   Supabase (Postgres) + Prisma ORM

Memory Architecture

The memory layer splits short-lived investigation state from long-term operational context. That keeps live investigations responsive while still letting the agent recall similar incidents and service relationships from older runs.

R
Redis
Working memory. Active investigation context, TTL-based.
Q
Qdrant
Long-term semantic memory. Past investigations by similarity.
N
Neo4j
Service dependency graph. Relationships between services and incidents.
S
Supabase
Audit log, user data, integration tokens, memory versioning.