Grove
Private memory across your AI agents, by Bramble Labs Inc.
Keep useful preferences, confirmed decisions, facts, verified solutions, and bounded
task handoffs. Browse, correct, archive, export, or delete them in your memory library.
Memory and notebook tools: recall, search, exact reads, partial edits, history,
filing, archive/restore and deletion. Tool reference.
One boundary: every website and agent request goes through the authenticated API.
One database: PostgreSQL full-text search, pg_trgm and pgvector, with account-owned rows
and non-owner runtime roles enforcing row-level security.
Website · Tools/API ·
Agent installation · Hosting ·
Delivery evidence
The website's /docs reference hub builds directly from selected committed docs.
See public reference maintenance for the
allowlist, link handling and privacy boundaries. Source delivery does not imply
that the new routes are already deployed.
Note search prioritizes visible titles, supports
final-word prefixes and conservative typo recovery, and shows source-backed
passages without requiring an AI provider.
Configurable embedding, decision and generation providers
add semantic search, generated titles/tags, related memories, and source-linked
personal profiles, project briefs and task handoffs. Cloud uses full 4,096-dimensional
Qwen embeddings and Clef decisions on DeepInfra, with MiniMax M3.1 Flash generation.
Connect a web chat
Use the remote MCP endpoint https://api.usegrove.cloud/mcp in ChatGPT or
Claude web, sign in with OAuth, then approve the matching client in Grove
Connections. No terminal or API key is needed. This is a custom-connection beta,
not a public-directory listing; host/account policies apply. See the
setup steps, permissions and verification limits.
Connect a coding agent
Connect to Grove Cloud, from your project directory (Node 22+ and your agent installed):
npx @miniaxolotl/grove connect
Grove detects your installed agent or lets you choose. You can also specify
connect opencode, connect claude or connect codex. Create an account or sign in
in the browser, then approve access. Grove links your current project, verifies
access and configures the integration. Keep the terminal open and follow its
agent-specific activation step.
Run the same command to resume interrupted setup.
Named, revocable API keys also support headless clients. Credentials stay in your
user configuration; plugins and hooks use the same connection. Ordinary agent
credentials cannot manage your account or issue other credentials.
Grove Open Source runs on your infrastructure. Add --url https://YOUR-GROVE-API
to connect to your own deployment. See the rollout requirements
for the release introducing connect; earlier CLI releases use grove init.
Host Grove
Requires Node 24/pnpm 10.33 for development, or Docker Compose for the service.
Configure a dedicated Grove Clerk application and website, then:
cp .env.example .env
# Set Clerk keys, webhook secret, and generated database/cursor passwords.
docker compose up -d --build
curl --fail http://localhost:26080/health/ready
Compose runs private pgvector PostgreSQL, forward-only migrations, a non-owner API,
and a separate job worker. The database has no published port. Local embeddings
initialize before readiness. See deployment for the
website, TLS, OAuth resource settings, backups, and Neon/DigitalOcean configuration.
Memory policy
- Durable saves remain until explicitly changed or deleted.
- Handoffs leave active recall after 30 days by default, with a 60-day restore window.
- Archive is reversible and excluded from recall. Deletion removes access and
derived work; authored notes and history remain retained behind soft-delete
policies, without an automatic physical-purge deadline. See
retained deletion.
- Saves report indexing separately; lexical retrieval remains available while
embeddings are pending. The embedding model never silently changes.
- Automatic recall is bounded; hooks never upload raw transcripts or tool logs.
- Server-side search is not end-to-end encryption. Recalled text has also been
delivered to your agent, whose retention policies are separate.
Development
pnpm install --frozen-lockfile
pnpm build
pnpm check-types
pnpm lint
pnpm dev:website
Documentation is the current topic index.
Verification commands cover actual
PostgreSQL/RLS, authenticated HTTP/MCP, packed stdio clients, and browser checks.
Reviewed file imports bring Markdown, plain text and
text-based PDFs into existing or new notebooks without overwriting existing notes.
Legacy import is explicit, ownership-mapped, and
source-read-only. Local execution plans are in
~/Development/bramble-labs/plans/grove/ (plan.md, notes.md, status.md);
recovered sources/ and the documentation archive
are historical evidence, not current implementation plans.
Current implementation contracts are documented under media,
recordings and speech;
production activation follows the release and explicit migration procedure.
Releases
Release Please manages one coordinated product version across the server, CLI
and three integrations. A released tag is deliberately frozen and promoted from
development through a protected PR to production. Only production publishes
packages, images and the website. See the release runbook.
License
MIT
