Sno Memory for Codex CLI

Close a Codex session and most of what it learned is gone. Sno Memory keeps it: an encrypted memory on your own machine that Codex fills as you work and reads from at the start of every session.

New in 2.0: Tag Team. Run sno reach pair "Fix the login redirect" and two agents take the task: one builds, one challenges, and the work reaches you already argued over. sno squad shows who needs you. An interrupted task comes back, or you are told it did not, and a lesson you approve works from the next reply. What changed.

What you get

• It remembers without being asked. Each session is captured in the background once it has been quiet for 30 minutes or has ended. Relevant memories come back when a session starts and before each prompt.
• Say it, and it is stored. remember saves a fact right now, recall looks one up, and get opens a full entry by id.
• Mistakes get corrected, not erased. correct writes the right text as a new entry and keeps the old one as retired history. Automatic recall only shows what is current.
• Scoped to your work. Memory belongs to the git repository you are in, plus a global scope you can read everywhere. Other agents that use the same setup (Claude Code, OpenClaw, Hermes Agent) use the same global scope.
• Yours, and it stays yours. The store is encrypted from first use with a key made once on your machine. Sno never receives your database or your key. There is no daemon to babysit: the memory service starts when Codex needs it.
• No account, no API key. The default mode runs on your machine. After a one-time model download during setup, recall works offline. It works in English, Chinese, Japanese, Korean, German, French, Spanish and Russian.

Get started

1. Follow the memory service setup. It takes a few minutes and runs once for every agent on the machine.

2. Install the plugin (Node.js 22.22.3+, 24.15.0+ or 25.9.0+, and git):

   curl -fsSL https://sno.ai/install | sh

3. Open a new Codex session inside a git repository, then try it:

   sno memory doctor --harness codex
   sno memory remember --harness codex "Prefer tabs in this repository"
   sno memory recall --harness codex "indentation"

   doctor should report the memory service as healthy (the line reads sidecar: healthy). To fix a wrong entry: sno memory correct --harness codex <id> "correct text". Child agents do not read or write memory.

Walkthrough: Codex onboarding. Daily use and troubleshooting: usage guide.

Modes

Local First (what the memory service setup selects) needs no model and no API key: capture is verbatim and everything stays on your machine. Agent Native lets Codex itself do the memory work on your existing subscription, and REM Enhanced adds Sno's own memory models to REM, the overnight step where your agents dream: it strengthens their memory and settles what they learned. It runs at most once a day for each part of your memory: usually overnight, or earlier the same day once about 100 new memories have piled up. Sno's models also write down what matters from each conversation once the session has been quiet for 30 minutes or has ended. It needs a Sno key in settings.json. In that mode your memory text and lessons go to your own space in Sno's cloud, which works like your own cloud drive: a Sno program processes it for you alone, and you can delete or keep it as you like. Change the mode in ~/.sno/settings.json, and keep the existing encryption key.

Under the hood

This package is the Codex side only: it installs Codex hooks and adds the memory commands. The memory itself lives in @snoai/memory, a local service on 127.0.0.1 that owns the encrypted SQLite store, vector and keyword search, and the local embedding model.

License

Licensed under Apache-2.0.