sno-ai/sno-station

Sno Station – a local “agent squad” workstation

What it is – Sno Station is an open‑source tool that lets you run multiple LLM‑based agents (Claude Code, Codex, OpenClaw) side‑by‑side on a single laptop. The agents share a encrypted local memory store and can hand‑off work when one hits its usage quota. A nightly RSI loop reads the agents’ session logs, spots recurring mistakes and proposes edits to the agents’ skill files; you approve those edits before they take effect.

Key features

• Dual‑agent “Duo”: two different model back‑ends work together, reviewing each other’s output to catch errors.
• Local‑only operation: no background daemon or cloud service is required; all memory and skill files stay on your machine and are encrypted with a user‑provided key.
• Multi‑language support: the workspace stores and classifies memories in English, Chinese, Japanese, Korean, German, French, Spanish and Russian.
• Memory that forgets: implements the Memora benchmark ideas, actively retiring stale facts and keeping the memory size bounded.
• Nightly self‑improvement loop: reads past sessions, measures impact of prior skill changes, and generates new skill‑file proposals.
• One‑command onboarding (planned): Sno onboarding inside any supported agent will install the CLI, shared memory, Duo skills and hooks.

How you use it

1. Install the (future) sno CLI via the onboarding command.
2. Run your usual agent (Claude Code, Codex, or OpenClaw). When its quota is near limit, invoke sno reach call from the other agent; it picks up the task with full context.
3. Ask an agent to peer‑review the other’s work for a second opinion.
4. Each night the RSI loop runs; in the morning you can accept suggested skill changes with rem‑reflect accept <id>.

Current status (as of 2026‑09‑20)

Component	Availability
packages/chunking	Published on npm
Core shared packages (common‑core, utils, …)	In repo
Cross‑model shared memory	Engine present, clean‑machine proof pending
Sno Reach (agent hand‑off)	Source present, release pending
RSI loop (self‑repair skill)	Shipping this week
One‑command install	Not yet released

Evidence of work – The README includes screenshots of the nightly RSI reports, quota‑watch hand‑off logs, and benchmark tables showing Sno Station’s performance on the Memora benchmark (FAMA 87.2, forgetting cost 3.76 pts) and other long‑memory tests.

License & security – Apache‑2.0. Memory is encrypted locally; the security model is described in docs/security.md.

Who might benefit – Developers who already use LLM agents for coding or automation and want a local, privacy‑preserving way to combine multiple agents, keep a persistent, self‑pruning memory, and have the system suggest skill improvements automatically.