Roverb

A personal AI memory you write to from Claude / Codex and recall anywhere — over MCP.

You're working in an AI tool, you hit something worth keeping, you say "remember this in Roverb." Weeks later, in any tool, you say "pull what I saved about X" and it comes back. One shared store behind all your assistants. No more scrolling old sessions.

• Local-first — one SQLite file at ~/.roverb/roverb.db. No cloud, no API keys, no telemetry.
• MCP-native — works with any Model Context Protocol client (Claude Code, Claude Desktop, Codex, Cursor, …).
• 7 tools — roverb_save, roverb_recall, roverb_get, roverb_list, roverb_update, roverb_forget, roverb_restore.
• Smart full-text search — SQLite FTS5 with BM25 ranking, prefix matching (so "rate" finds "rate-limiting"), a recency nudge, and highlighted snippets that show why each result matched. (Semantic/vector recall is a later add-on.)
• Safe by default — forget moves a memory to the trash (recoverable); nothing is permanently deleted unless you purge.
• Backup & portable — roverb export dumps everything to JSON, roverb import loads it back (idempotent, so re-imports won't duplicate).

What you actually run

Roverb has two pieces, both on the same local database:

| Piece | What it's for | How it runs | |---|---|---| | MCP server | the agent side — Claude/Codex call it to save & recall | started automatically by your AI tool after roverb init (you don't launch it by hand) | | Dashboard | the browser UI — see, search, save, forget | you run it yourself: roverb ui → http://localhost:4319 |

So after setup there's really one command to remember: roverb ui to open the dashboard. The MCP server takes care of itself.

Quick start (the easy way)

Requires Node 18+ — nothing to install or clone:

# 1. point your AI tools at Roverb (auto-configures Codex, Claude, Cursor)
npx -y roverb init

# 2. open the dashboard in your browser (optional, but nice)
npx -y roverb ui      # → http://localhost:4319

Then restart your AI tool and say "remember this in Roverb." That's it — init already configured the MCP server, and your tool launches it on demand.

roverb init writes the MCP config for whatever it finds installed:

• Codex → ~/.codex/config.toml
• Claude Code → via claude mcp add -s user (user scope → available in every project, not just the folder you ran it in)
• Claude Desktop → claude_desktop_config.json
• Cursor → ~/.cursor/mcp.json

All of them are registered as npx -y roverb@latest mcp. Your tool launches the server via npx on every start, which contacts the npm registry — so @latest simply guarantees you get the newest published version each launch, and updates land automatically with no re-config.

Faster/offline setup: npx contacts the npm registry when the server starts. If you want the fastest startup, or you work offline/behind a proxy, install globally and initialize the global command instead:

npm i -g roverb
roverb init --global --force

That registers command: "roverb", args: ["mcp"] with a short Codex startup timeout. Updates are then manual via npm update -g roverb.

Each person who runs roverb init gets their own local store at ~/.roverb/roverb.db — memories stay on your machine and aren't shared between people.

Prefer not to use the npm release?

Run it straight from GitHub instead — still no clone or build:

npx -y github:hakobyan-arsen/roverb init
npx -y github:hakobyan-arsen/roverb ui

See it in a browser (local dashboard)

npx -y roverb ui       # → http://localhost:4319

Four screens: Ask (search/recall), Library (browse + filter, forget, plus a Trash filter to restore or empty forgotten memories), Capture (save), and Connection (stats, backup export/import, MCP config). Small REST API in src/server.js, no extra deps. Override the port with ROVERB_PORT. The dashboard and the MCP server share the one database, so run both at once.

Manual config (if you skip init)

The command to register is npx -y roverb@latest mcp.

Codex — ~/.codex/config.toml:

[mcp_servers.roverb]
command = "npx"
args = ["-y", "roverb@latest", "mcp"]
enabled = true
startup_timeout_sec = 120

Claude Code (CLI) — user scope, so it's available in every project:

claude mcp add -s user roverb -- npx -y roverb@latest mcp

Claude Desktop / Cursor — mcpServers block:

{ "mcpServers": { "roverb": { "command": "npx", "args": ["-y", "roverb@latest", "mcp"] } } }

Restart the tool; you should see 5 roverb_* tools.

If you installed globally with npm i -g roverb, use the faster command instead:

Codex — ~/.codex/config.toml:

[mcp_servers.roverb]
command = "roverb"
args = ["mcp"]
enabled = true
startup_timeout_sec = 10

Or let Roverb write it:

roverb init --global --force

Use it

Just talk to your assistant — it calls the tools for you:

• "Remember this decision in Roverb, tag it api and infra." → roverb_save
• "What did I save about rate limiting?" → roverb_recall
• "List everything from the Billing project." → roverb_list
• "Forget that staging note." → roverb_recall to find it, then roverb_forget (it goes to the trash)
• "Actually, bring that note back." → roverb_restore

Make it proactive (optional)

Drop this in your project's CLAUDE.md / AGENTS.md so the agent reaches for Roverb without being told:

You have a `roverb_*` memory tool set (an MCP server called "roverb").
- Before answering questions about past decisions, preferences, or project
  history, call roverb_recall first.
- When the user makes a decision, states a preference, or produces something
  reusable, offer to roverb_save it.

A memory record

{
  "id": 12,
  "type": "decision",          // decision | code | note | link | fact | writing | image
  "title": "Postgres over Mongo for billing",
  "body": "Billing needs ACID guarantees ...",
  "source": "claude",          // which tool it came from
  "project": "Billing",
  "tags": ["db", "billing"],
  "created_at": "2026-06-22T10:00:00.000Z",
  "updated_at": "2026-06-22T10:00:00.000Z",
  "archived_at": null,         // set when forgotten (in the trash); null = active
  "last_accessed": null,       // last time fetched via roverb_get
  "access_count": 0            // how many times it's been fetched
}

Storage path override: set ROVERB_STORE=/some/path/roverb.db.

Back up & restore

roverb export --out roverb-backup.json   # dump everything (including trash) to JSON
roverb import roverb-backup.json          # load it back — safe to re-run, duplicates are skipped

License

MIT.