Sno Station 1.6.0: a local model now reranks memory search, in every mode

Sno Station 1.6.0 changes how memory search orders its results. In every memory mode, a small ranking model on your own computer now reorders what search finds, so your prompts and memory text stay local while it ranks.

Slide text

Sno Station 1.6.0 sno-ai/sno-station

A local model now reranks memory search, in every mode

Prompts and memory text stay local while results are reordered.

Search still starts the way it did, retrieving memories and putting them in a first order on your machine. Then a small multilingual ranking model, running on your CPU, reads your query beside each of the first 32 candidates and reorders them. The usual scoring and your result limit come after, and importance weights no longer undo the model's order.

Slide text

Sno Station 1.6.0 Added

Search finds candidates, then a local model reorders the first 32

your machine

Retrieve and order Memories found and ordered locally, as before

→

Local model reranks Reads the query beside each of the first 32 candidates multilingual-reranker-small · 8-bit · CPU

→

Final scoring Importance and diversity weights, then your result limit

The onboarding guide now gives ranking its own row in the mode table. It runs on your machine in Local First, Agent Native and REM Enhanced alike, and reranking sends no prompt or memory text to Sno GPU. Memory writing, REM and Self Learning still go wherever each mode sends them.

Slide text

Sno Station 1.6.0 Privacy

Ranking stays on your machine in every mode

What it does	Local First	Agent Native	REM Enhanced
Writing down what matters	Saved word for word	Your Claude Code	Sno's memory model
Settling memories that disagree	Keep both	Your Claude Code	Sno's memory model
Ranking what to recall	On your machine	On your machine	On your machine

Memory writing, REM and Self Learning keep their mode-dependent destinations.

docs/mem-claude/onboarding.md · mode-specific defaults, excerpt

Search answers without the model when it has to. Your first recall starts a download of about 206 megabytes in the background, and until it's ready, results keep their first order. After that, the model gets two seconds at most, and automatic recall gives it no more than half of its own time limit. Late scores are dropped. A model that failed to load is tried again on the first recall after ten minutes.

Slide text

Sno Station 1.6.0 Fallback

Search returns results while the model loads, fails or runs late

first recall

The model, about 206 MB, downloads in the background. Results keep their first order.

2 s at most

Each rerank gets two seconds, and auto-recall gives it at most half its own time limit. Late scores are dropped.

after 10 min

A model that failed to load is tried again on the next recall.

Both models now arrive through one shared downloader. A broken connection or a killed process picks up from the bytes already on disk. Downloads use the proxy your environment names. A connection that goes quiet for 30 seconds is retried, instead of hanging for five minutes. And every file is checked against a pinned hash.

Slide text

Sno Station 1.6.0 · embedder 1.2.1 Fixed

Model downloads resume, retry and use your proxy

Resumes From the bytes already on disk, after a broken connection or a killed process

Uses your proxy HTTP_PROXY · HTTPS_PROXY · NO_PROXY

Drops a stalled connection After 30 seconds of silence, then retries, instead of waiting five minutes

Checks every file Against its pinned SHA-256 hash, and fetches a mismatch again

Each agent's setup guide got shorter. The Claude Code guide used to run sno setup after the installer, with a flag naming the agent. Now the one command sets up Sno Station and connects every agent it finds. If you add an agent later, run sno setup again.

Slide text

Sno Station 1.6.0 Docs

The install command sets up Sno Station on its own

Claude Code guide, before

sh -c 'sno_installer_body=$(curl -fsSL https://sno.ai/install) && printf "%s\n" "$sno_installer_body" | sh' && ~/.local/bin/sno setup --harness claude

Now

sh -c 'sno_installer_body=$(curl -fsSL https://sno.ai/install) && printf "%s\n" "$sno_installer_body" | sh'

Installed another agent later? Run sno setup again.

In 1.3.5 you could send reranking to a remote provider, such as Voyage, Jina or your own server, with a provider, an endpoint, a model and a key. Those fields are gone. A mode and a time limit remain. A saved cross encoder mode now means the local model, and the time limit is capped at two seconds. Set the mode to none to skip the model.

Slide text

Sno Station 1.6.0 · settings.json Removed

Remote reranker settings are gone; two fields remain

1.3.5 default

"rerank": {
  "mode": "lightweight",
  "provider": "voyage",
  "endpoint": "",
  "model": "rerank-2",
  "apiKey": "",
  "timeoutMs": 15000,
  "maxCandidates": 64
}

1.6.0 default

"rerank": {
  "mode": "lightweight",
  "timeoutMs": 2000
}

A saved cross-encoder is read as lightweight, the local model.

none keeps the first order, without the model.

To upgrade, run sno update. If you set up the memory service by hand, install memory 1.6.0 with embedder 1.2.1, reinstall your agent plugin, and restart running sessions, without rerunning the settings block. Hold off only if you depend on a remote reranker you configured. Otherwise, from 1.6.0 on, a local model reranks memory search, in every mode.

Slide text

Sno Station 1.6.0 Upgrade

Upgrade with one command

sno update

Memory service set up by hand

npm install --prefix "$HOME/.sno" @snoai/memory@1.6.0 @snoai/embedder@1.2.1

Then reinstall your agent plugin and restart running sessions. Don't rerun the settings block.

Wait if your recall relies on a remote reranker you configured. In 1.6.0 the local model ranks instead.