Local memory bridge · laptop setup

Install Velvet Signal locally.

Give an Ollama model a durable, user-controlled shelf of approved context patches. The model itself does not need persistent memory: Velvet Signal stores the patches, retrieves what matters, and supplies the current claims at inference time.

● Your local questions stay with your configured Ollama server

Step 01

Install the three tools.

You need Git, Node.js 22 or newer, and Ollama. Git downloads the project, Node runs Velvet Signal, and Ollama runs the local chat and embedding models.

Git
Node.js 22+
Ollama

On Windows, Node can be installed from PowerShell:

winget install OpenJS.NodeJS.LTS

Close and reopen PowerShell, then verify:

node -v
npm.cmd -v
git --version
ollama --version

If PowerShell says npm.ps1 cannot be loaded because running scripts is disabled, nothing is wrong with Node. Use npm.cmd instead of npm. You do not need to loosen PowerShell security settings.

Step 02

Download Velvet Signal.

Start from your own user folder—not System32—and clone the repository.

cd $HOME
git clone https://github.com/TONZHub/Velvet_Signal.git
cd Velvet_Signal

If you are testing a feature branch before it reaches main, switch to that branch only when the project documentation tells you to.

Step 03

Install the Node dependencies.

Run this while your terminal is inside the Velvet_Signal folder.

npm.cmd install

macOS/Linux users can use npm install. If npm says it cannot find package.json, you are in the wrong folder. On Windows, run cd $HOME\Velvet_Signal first.

Step 04

Prepare Ollama.

Velvet Signal uses a small local embedding model to decide which claims are relevant. Then it can send the selected context to whichever Ollama chat model you choose.

ollama pull embeddinggemma
ollama list

If you need a chat model, pull one. For example:

ollama pull dolphin3:8b

Step 05

Let a local model call the live tools.

Velvet Signal exposes the server-side counterparts of its four browser WebMCP tools through a standard MCP Streamable HTTP endpoint at https://velvetsignal.lol/mcp. Use a tool-capable Ollama model to discover that endpoint and decide which tools to call.

ollama pull qwen3:4b
npm.cmd run webmcp:demo -- --model qwen3:4b

The terminal prints each boundary crossing: the tools discovered from the site, each call selected by the local model, and the result returned to it. The default proof asks the model to list Maker Edition issues and inspect maker-012.

This is an explicit WebMCP-to-MCP bridge, not a claim that Ollama can reach into Chrome. The browser and HTTP endpoint expose the same named publication tools. Browser approvals stay in browser storage; the HTTP bridge can inspect a patch but cannot approve one or mint a receipt.

Step 06

Release a patch into local memory.

Velvet Signal does not silently write publication patches into your local ledger. You choose which patch enters it.

npm.cmd run local -- release pantry-003
npm.cmd run local -- list

Released patches live in ~/.velvet-signal/patches.json, outside the repository. That ledger survives terminal, Ollama, code-update, and computer restarts.

Step 07

See what Velvet Signal remembers.

Inspect retrieval before asking a language model. This lets you separate a retrieval problem from a model reasoning problem.

npm.cmd run local -- inspect "I cooked chicken five days ago and kept it refrigerated. It smells fine. Can I eat it?"

You should see ranked claims with patch IDs, claim IDs, source IDs, validity dates, and semantic/lexical scores.

Step 08

Ask your local model.

Replace the model name with anything shown by ollama list.

npm.cmd run local -- ask --model dolphin3:8b "I cooked chicken five days ago and kept it refrigerated. It smells fine. Can I eat it?"

Velvet Signal retrieves a small set of active claims, keeps their provenance attached, and supplies that context to the current user turn. The local model is free to reason over it; Velvet Signal does not rewrite model weights.

Later

Update the code without erasing memory.

cd $HOME\Velvet_Signal
git pull
npm.cmd install

Your patch ledger is stored outside the repository, so pulling new code does not delete previously released local patches.

Windows troubleshooting

The errors we already stepped on.

npm : The term 'npm' is not recognized

Install Node.js, close PowerShell completely, reopen it, then try node -v and npm.cmd -v.

npm.ps1 cannot be loaded because running scripts is disabled

Use npm.cmd. No execution-policy change is required.

EPERM ... C:\Windows\System32\package-lock.json

You are running npm from System32. Do not run as Administrator to force it. Run cd $HOME\Velvet_Signal and try again.

ENOENT ... C:\Users\you\package.json

You are in your home folder instead of the repository. Run cd $HOME\Velvet_Signal.

Ollama cannot find the model

Run ollama list, then ollama pull <model-name> for the chat or embedding model you want.