v3.2 Runtime Bridge shipped — RAG Runtime Kernel is live

[arcadamarket]

The v3.2 Runtime Bridge is now live.

What is RAG Runtime Kernel?

A filesystem-backed, event-sourced memory system for LLMs. One markdown specification file gives any LLM persistent, crash-recoverable state management.

Core principle: LLM proposes. System decides. State persists.

What shipped in v3.2

• 8 Python modules (state machine, persistence, COLD manager, concurrency, API, MCP transport, CLI, schemas)
• 337 passing tests, 5811 lines of enforcement runtime
• ENFORCED mode: hard runtime validation of every state transition
• The spec itself is 43 sections (v3.1.6)

Two operational modes

• AUTONOMOUS — paste the markdown spec into any LLM context. Zero code, zero install.
• ENFORCED — Python runtime validates every proposal against schema and transition rules.

Key architecture

• Deterministic state machine: BOOTING → READY → WORKING → CHECKPOINTING → CLOSING
• Proposal → Validate → Commit contract (borrowed from database transaction theory)
• Write-ahead logging + atomic writes + hash verification
• HOT/COLD memory partitioning for context window efficiency
• LLM-agnostic: works with Claude, GPT, Llama, Mistral, or any model

Get started

Clone the repo, open the spec file, paste it into your LLM's context. Start working. State persists between sessions.

AGPL-3.0 licensed. No telemetry, no accounts, no vendor lock-in.

Feedback, questions, and PRs welcome.