sno-ai/sno-station is a locally running multi-agent harness that bundles Codex, Claude Code, Cursor, Hermes Agent, and OpenClaw into a single encrypted memory, hands work between them, and has them review each other. The latest release is v1.3.9, written in TypeScript under the Apache-2.0 license. The repository shows 758 stars.
When one agent hits a usage limit or its context window, the next agent picks up the work from the same memory. The key is carrying prior context forward even when the session changes. The project explains that the reviewer is fixed to a different agent to prevent the tendency to unconditionally agree with the first agent. The nightly REM loop reviews the session once a day at most, or when roughly 100 new memories have piled up, looking for recurring mistakes and proposing skill changes. Proposed changes are not applied until the user approves them. In local-first mode it does not upload memory bodies, but in REM Enhanced mode the bodies and some of the learned proposals are uploaded to user space. It emphasizes operations that run to completion locally with no daemon and no account.
The LongMemEval-S retrieval R@5 of 95.4% cited as evidence of memory quality is the project's own benchmark result. With no public replay harness, it is hard to reproduce and verify externally. In REM Enhanced mode, prompts and up to 64 stored memories are sent to an external cloud, and the defaults are stated to upload a day's session to the cloud. So the emphasis on local execution alone makes it difficult to judge the data path and the memory quality.