omni-memory

A personal, fully-local AI memory system. Save notes, code snippets, or terminal output to a vector database on your machine, then retrieve relevant context from the terminal, direct API calls, or automatically inside Claude Code.

No cloud, no external services - embeddings run locally via Ollama and vectors are stored in a single SQLite file with the sqlite-vec extension.

Architecture

remember "..."  ┐
ingest scripts  ┼──► Express hub (127.0.0.1:8000) ──► SQLite + sqlite-vec (memories.db)
Claude Code MCP ┘            │
                             └──► Ollama (nomic-embed-text, 768-dim embeddings)

• src/server.ts - Express hub; API-key auth + centralized error handling.

• src/routes/ - remember, query, list, forget, health.

• src/lib/ - db (schema + all data access: insert, cosine search, list, delete), embed, chunk, dedupe, store (shared insert path), sql (parameterized filter builders), types.

• src/mcp/index.ts — MCP server exposing search_memory, save_memory, forget_memory.

• scripts/ — ingest, ingest-folder, query, dedupe-existing CLIs.

Related MCP server: CozoDB Memory MCP Server

Install on a new machine

Prerequisites

• Node.js 20+ and npm — https://nodejs.org

  • The storage layer (better-sqlite3, sqlite-vec) installs prebuilt native binaries on Windows x64, macOS, and Linux x64/arm64 - no compiler needed. Windows on ARM is not covered by sqlite-vec's prebuilds.

• Git

• Ollama — install from https://ollama.com/download (Windows/macOS/Linux), then confirm it's running: ollama --version. It serves on http://localhost:11434.

• Claude Code (optional, for the MCP tools) — https://claude.com/claude-code

1. Clone and install

git clone <your-repo-url> omni-memory
cd omni-memory
npm install

2. Pull the embedding model

ollama pull nomic-embed-text

3. Configure .env

Create .env in the project root (it is gitignored — never commit real keys). Generate a random API key with Node:

node -e "console.log('omni-' + require('crypto').randomBytes(16).toString('hex'))"

Then put it in .env, and point DB_PATH at where you want vectors stored (use a path for your OS — any writable folder, created on first use):

OMNI_API_KEY=omni-<the-generated-key>
DB_PATH=C:/Users/<you>/.ai_memory      # Windows
# DB_PATH=/home/<you>/.ai_memory       # Linux
# DB_PATH=/Users/<you>/.ai_memory      # macOS
OLLAMA_URL=http://localhost:11434
PORT=8000

If DB_PATH is omitted it defaults to ~/.ai_memory (via USERPROFILE/HOME).

4. Build

npm run build

5. Start the hub

Windows — register a hidden Task Scheduler task that starts at every logon:

.\install-service.ps1

macOS / Linux — run in the foreground (any OS):

npm start            # or `npm run dev` for tsx without a build

To start it automatically at login on Linux with systemd, create a user service at ~/.config/systemd/user/omni-memory.service:

[Unit]
Description=omni-memory hub
After=network.target

[Service]
WorkingDirectory=%h/omni-memory
ExecStart=/usr/bin/node %h/omni-memory/dist/server.js
Restart=on-failure

[Install]
WantedBy=default.target

Then enable it (adjust the paths to your clone and node):

systemctl --user daemon-reload
systemctl --user enable --now omni-memory
systemctl --user status omni-memory     # check it's running

On macOS, the simplest equivalent is a launchd agent in ~/Library/LaunchAgents/, or just run npm start in a terminal.

The hub creates the SQLite database (memories.db inside DB_PATH) on first start. Backup is just copying that file.

6. Register the MCP server in Claude Code

So Claude can call search_memory / save_memory / forget_memory, register the server at user scope (use the absolute path to your clone's dist/mcp/index.js, and the same key as .env):

claude mcp add omni-memory --scope user \
  --env OMNI_API_KEY=omni-<the-generated-key> \
  --env OMNI_HUB_URL=http://127.0.0.1:8000 \
  -- node /path/to/omni-memory/dist/mcp/index.js

This writes the entry into ~/.claude.json. To configure it by hand instead, add this under mcpServers in ~/.claude.json:

{
  "mcpServers": {
    "omni-memory": {
      "command": "node",
      "args": ["C:\\path\\to\\omni-memory\\dist\\mcp\\index.js"],
      "env": {
        "OMNI_API_KEY": "omni-<the-generated-key>",
        "OMNI_HUB_URL": "http://127.0.0.1:8000"
      }
    }
  }
}

Use the right path style for your OS in args:

• Windows: "C:\\path\\to\\omni-memory\\dist\\mcp\\index.js" (escaped backslashes)

• Linux/macOS: "/home/<you>/omni-memory/dist/mcp/index.js"

Restart Claude Code (MCP config is read at startup) with the hub already running.

7. (Optional) remember shortcut

So you can save from any terminal with remember some note:

Windows — add to your $PROFILE:

$env:OMNI_KEY = "omni-<the-generated-key>"     # same key as .env
$script:_omniHub = "http://127.0.0.1:8000"
function remember {
    param([Parameter(ValueFromRemainingArguments)][string[]]$words)
    $body = @{ text = ($words -join ' '); source_type = "terminal" } | ConvertTo-Json
    Invoke-RestMethod -Uri "$script:_omniHub/remember" -Method POST `
        -Body $body -ContentType "application/json" `
        -Headers @{ "X-API-Key" = $env:OMNI_KEY }
}

Reload with . $PROFILE.

Linux / macOS — add to your ~/.bashrc or ~/.zshrc:

export OMNI_KEY="omni-<the-generated-key>"     # same key as .env
remember() {
    local body
    body=$(python3 -c 'import json,sys; print(json.dumps({"text": " ".join(sys.argv[1:]), "source_type": "terminal"}))' "$@")
    curl -s -X POST http://127.0.0.1:8000/remember \
        -H "X-API-Key: $OMNI_KEY" -H "Content-Type: application/json" \
        -d "$body"
    echo
}

Reload with source ~/.bashrc (or ~/.zshrc). Uses python3 for JSON escaping so it needs no extra tools; if you prefer jq: -d "$(jq -n --arg t "$*" '{text:$t, source_type:"terminal"}')".

8. Verify

# PowerShell
irm http://127.0.0.1:8000/health -Headers @{"X-API-Key"="omni-<key>"}
# bash
curl -H "X-API-Key: omni-<key>" http://127.0.0.1:8000/health

Expect {"status":"ok","ollama":"ok","count":0}. In Claude Code, ask it to "search my memory for X" to confirm the MCP tools are wired up.

API

All endpoints require header X-API-Key: <key>. Hub listens on 127.0.0.1:8000.

• source_type / source: terminal | chat | code.

• importance: 0–1 (default 0.5); higher ranks earlier in /query results.

• tags: matches memories containing any of the given tags.

• /remember returns 400 when no chunk longer than 20 characters survives chunking — write at least a full sentence.

Development

npm run dev    # run the hub with tsx (no build step)
npm test       # run the Vitest unit suite
npm run build  # compile to dist/

After changing src/, rebuild and restart the service cleanly — kill any stray server.js process first, or the old code keeps port 8000 (see HOW-TO-USE.md).

See HOW-TO-USE.md for daily usage, ingest recipes, and troubleshooting.