sno-ai/sno-station

What it claims

Sno Station is Apache-2.0 software that pools the AI coding/working agents you already run (it names Claude Code, Codex, and OpenClaw) into one "Duo" squad on your own machine, with no daemon, server, or cloud required. Three behaviors are the pitch: when one agent hits its rate limit, another from a different harness resumes the task with context intact via sno reach call; either agent can peer-review the other's work (the reviewer is always from the other harness); and a nightly "RSI loop" reads the agents' sessions, finds repeated mistakes, and proposes edits to the agents' own skill files — applied only when a human runs rem-reflect accept <id>.

Critically, the repo is being opened "one piece at a time" starting 2026-09-18. The one-command install does not exist yet ("Not yet," last updated 2026-09-19); the RSI loop was slated to ship "this week." What is present is documentation, evidence logs, and eval records — not yet a runnable install.

How it works

Memory is the stated foundation. It stores each memory with its language, classifies by locale, preserves CJK text in keys and search (nine languages), and is designed to forget on purpose — retiring cancelled/superseded facts, not just recalling. The store is encrypted on-device with a key provisioned once that never leaves the machine; Sno claims it never receives your database or key (boundary in docs/security.md). Design is credited to Karpathy's "LLM Wiki," Google/Virginia Tech's WikiSkill (arXiv:2608.27454), and Stanford/SambaNova's Agentic Context Engineering (arXiv:2510.04618).

Evidence

Self-run result on the Memora forgetting benchmark (Uddin et al., arXiv:2604.20006, ACL 2026 Findings), weekly track: six personas, ninety questions, one formal run dated 2026-09-06. Reported reasoning (multi-memory combination) of 93.3 against a published field of 2–30/100, and FAA (share of cancelled facts correctly excluded) of 0.833. The authors flag this as directional, not certified: the paper's six agents were judged with GPT-4o-mini, Sno's with its own judge stack recomputed under the paper's penalty formula. The paper publishes no per-agent FAA, so no competitor forgetting comparison is shown. Also cited: a 2026-09-18 handover log (docs/evidence/rotation-2026-09-18/) where an agent at 2% quota briefed a second harness and handed off in 14m41s.

Limits

• No install exists yet; near-everything actionable is "watch this repository."
• The benchmark comparison is self-reported and self-judged, not a certified head-to-head.
• Cloud features are "when it comes" — unbuilt.
• Verify the encryption/trust boundary claims in docs/security.md before relying on them.

For an engineer: nothing to adopt today, but worth tracking if you run two+ agents by hand and want cross-harness handover, review, and forgetting-aware memory.