Sno AI released sno-station, an Apache-2 local tool that lets Claude Code and Codex share one encrypted memory and review each other's work, with 402 GitHub stars in two weeks: Reveneau

Why it mattersA developer paying for two coding agents today runs each one in its own window with its own memory, so a cross-check happens in a human's head; sno-station moves that check into the tooling and keeps the context when one agent runs out of quota.

A developer who pays for both Claude Code and Codex runs each one in its own terminal, with its own memory, and copies context across by hand when one agent hits a rate limit. Sno AI released sno-station under Apache-2.0, giving both agents one encrypted memory on the laptop they already run on, a direct messaging channel between them, and a review step where one checks the other's work before it reaches the developer. The repository reports 402 stars and 322 forks on 4 October 2026, per the GitHub REST API, with the first commit on 19 September 2026.

What ships in the box

The README names four agents the station already coordinates: Claude Code, Codex, OpenClaw and Hermes Agent. All four read and write to the same local store, which the README describes as "no daemon, no server, no cloud required; the cloud side, when it comes, is optional and the product is complete without it." That is the whole operating model: a shared folder on your machine, not a service you log into.

Three mechanisms sit on top of that store. Sno Reach is the message channel between agents, so one can hand a task to another with the full working context attached. Squad skills are shared instructions written once and read by every agent in the station, so a handoff carries the same rules the first agent was following. A nightly loop analyses the day's sessions and proposes rewrites to those skills, which the README states "rewrites the agents' own skills with your approval"; the human stays in the gate on every change.

Memory, with the numbers the project reports

Sno AI reports its own scores on four memory benchmarks, which the README states and which belong to the project rather than to an independent run. On the Memora benchmark, the station reports a Forgetting-Aware Memory Accuracy of 87.2, with a forgetting cost of 3.76 points, or 4.1 percent of recall. On LoCoMo, a long-conversation question set, the project reports 96.76 percent accuracy over 1,542 questions. On LongMemEval-S, which measures whether the right memory is retrieved before the agent answers, the project reports 95.4 percent recall in the top five results. Sno Memory Bench, which the project built and runs itself, reports 23 of 23 passing end-to-end probes. Treat these as the vendor's own numbers; the Memora, LoCoMo and LongMemEval teams did not publish these runs.

Cross-vendor review is where the design pays off

The most practical piece for a working team sits in the peer-review step. When one agent completes a task, a second agent from a different vendor reads the diff, the test output and the shared memory, and either passes it or writes a review that the first agent then has to answer. A developer running Claude Code and Codex side by side already does this review in their head when the two give different answers; sno-station moves the check into the tool, so a tired human catches fewer mistakes by not being in the loop.

The repository is young and the mechanisms are new. 402 stars in two weeks is real traction for a developer tool, and the Apache-2.0 licence means the code is yours to read and change.

Source

Sno Station GitHub repository, primary source: https://github.com/sno-ai/sno-station
GitHub REST API for star count, forks and dates, queried 4 October 2026: https://api.github.com/repos/sno-ai/sno-station

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.