GitHub Repository

The agent infrastructure that chatbot wrappers forgot to build. Missions, councils, sandboxing, hybrid memory, and human-in-the-loop controls. Not a wrapper — an OS. Apache 2.0.

14 starsTypeScript

Sovereign – Multi-agent OS with GraphRAG memory and HITL checkpoints

Name: Sovereign – Multi-agent OS with GraphRAG memory and HITL checkpoints
Availability: InStock
Author: borhensaidi

by borhensaidi·Feb 18, 2026·1 point·0 comments

Visit Project View on HN

AI Analysis

●●SolidWizardryNiche Gem

The Take

Runtime pause-and-resume checkpoints during execution plus per-agent escalation policies are a rare, practical safety feature — not just a pre-flight approval checkbox. The hybrid memory that scores vector similarity, BM25 keyword relevance, and entity/relation overlap together is a smart middle path between heavyweight graph DBs and plain vector stores. There's real engineering here (Docker compose, 55/55 tests, telemetry endpoints), though I'd like to see benchmarks and how the in-process GraphRAG scales versus external vector/graph services.

Post Description

I built this because existing agent frameworks either give agents god-mode filesystem access with no safety controls, or they're too locked down to be useful.

Key things I cared about that most frameworks skip:

- Runtime HITL checkpoints — pause mid-execution when risk is high, resume from exact state - Hybrid memory: vector + BM25 keyword + graph entities/relations scored together, no external dependencies - Security as a primitive: path jail sandbox, OTP pairing, XOR-encrypted local secrets - 22+ LLM providers with per-agent model policy and fallback chains - Multi-agent councils with debate rounds, soul evolution, and skill memory

Node.js, Prisma, Postgres/Redis, Docker. 55/55 tests passing. Apache 2.0.

Happy to answer questions about any of the architecture decisions.