Community code plugin. Review compatibility and verification before install.
Latest release: v0.7.69Download zip
Capabilities
Compatibility
Security Scan
OpenClaw
Suspicious
high confidencePurpose & Capability
Cloud-backed shared memory, auto-capture, and team recall are coherent with the stated purpose and disclosed in the README, but automatic skill mining, delegated agent execution, and global skill writes expand the capability beyond ordinary memory storage.
Instruction Scope
The injected agent context encourages use of org-shared summaries and raw session data across teammates, and the skill-mining documentation in the embedded prompt conflicts with code that auto-spawns a mining worker after captured sessions.
Install Mechanism
The npm package does not show a postinstall-style execution path, and OpenClaw config edits are tied to explicit commands with backups; however those commands rewrite allowlists and update settings without an interactive diff or confirmation inside the code path.
Credentials
Sending every user and assistant message to api.deeplake.ai is purpose-aligned and README-disclosed, but it is broad, default-on after authentication, team-visible, and lacks evident redaction or per-message controls.
Persistence & Privilege
The plugin stores a long-lived Deeplake token, modifies OpenClaw config, spawns a detached skill-mining worker, writes local state/logs, stores derived skills remotely, and can write generated SKILL.md files under the user's global Claude skills directory.
Scan Findings in Context
[VirusTotal] expected: VirusTotal telemetry is clean and does not support a malicious verdict.
[staticScan] expected: The static scan reported clean; artifact review still found high-impact behavior that warrants Review rather than automatic release.
[SDI-2] expected: The allowlist and autoUpdate config writes are tied to documented slash commands and backups, so they are not hidden install-time changes, but they still alter trust and update policy with coarse user control.
[SDI-1] unexpected: The code auto-spawns skill mining after captured sessions, while embedded guidance says OpenClaw itself does not run sessions to mine; this mismatch is a material disclosure problem.
[SDI-4] unexpected: The session-mining contradiction is supported by artifact code and materially affects user expectations about secondary processing of conversations.
[SDI-2] unexpected: The skillify worker delegates mined session content to local agent CLIs using bypassPermissions, dangerously-bypass-approvals-and-sandbox, or yolo-style flags; this is high-impact and under-scoped for automatic background processing.
[SDI-4] unexpected: The scanner's specific global-read-only merge claim is only partially supported, but the worker is configured for global installs and can create or merge persistent SKILL.md files under the user's global skills path.
[SQP-2] expected: Remote auto-capture of conversations is disclosed in the README and matches the memory purpose, but the breadth, default-on behavior after login, and team visibility keep it a Review concern.
[SQP-2] unexpected: Remote storage of derived skill content and SQL queries from mined sessions is real, but the skill-mining data flow is less clearly disclosed than the core memory capture flow.
[SQP-1] expected: The manifest mentions auto-capture, auto-recall, and teammates, but the README carries most of the actual privacy disclosure; users relying only on install metadata would not see enough scope detail.
[SSD-3] expected: The auth URL exposure is part of the login flow, but placing a live link into model-visible context increases leakage risk through transcripts or logs.
[SSD-3] expected: Org-shared memory access is the advertised product behavior, but raw cross-user session search/read needs clear administrative access controls and user expectations.
What to consider before installing
Install only if you are comfortable with Deeplake receiving and retaining agent conversations for your org, including possible teammate access. Before enabling it, confirm capture can be paused or disabled, decide whether global skill mining is acceptable, and review the effects of /hivemind_setup, /hivemind_autoupdate, and any generated skills under ~/.claude/skills.Verification
Tags
Hivemind
Cloud-backed shared memory for AI agents. Install once, memory persists across sessions, machines, and channels — and is shared with every teammate in your Deeplake org.
Powered by Deeplake.
Install
openclaw plugins install clawhub:hivemind
Then in chat:
/hivemind_login
Click the auth link, sign in, send another message. That's it.
What it does
- Auto-recall — before each agent turn, relevant memories surface automatically via keyword search.
- Auto-capture — after each turn, the conversation is stored to your Deeplake workspace.
- Cross-platform — same memory accessible from Claude Code, Codex CLI, and OpenClaw plugins.
- Team-wide — every user in your Deeplake org shares the same memory.
Commands
| Command | What it does |
|---|---|
/hivemind_login | Sign in via device flow |
/hivemind_setup | Add hivemind to OpenClaw's tool allowlist (one-time, after install) |
/hivemind_capture | Toggle conversation capture on/off |
/hivemind_whoami | Show current org and workspace |
/hivemind_orgs | List organizations |
/hivemind_switch_org <name> | Switch organization |
/hivemind_workspaces | List workspaces |
/hivemind_switch_workspace <id> | Switch workspace |
/hivemind_version | Show installed version and check ClawHub for a newer one |
/hivemind_update | Show how to install the latest version |
/hivemind_autoupdate [on|off] | Toggle the agent-facing update nudge (on by default) |
You can also just ask the agent naturally — "switch org to activeloop", "list my orgs", "invite alice@example.com as admin", etc.
Privacy & data
- What's captured: every user message and assistant reply, sent to
api.deeplake.ai. - Where credentials live: a long-lived API token at
~/.deeplake/credentials.json(file permissions 0600). - Where it sends data:
api.deeplake.ai(memory storage) andclawhub.ai(version check on session start and via/hivemind_version). - How to pause: run
/hivemind_captureto stop capture; run it again to resume. - How to fully sign out: delete
~/.deeplake/credentials.jsonand revoke the token in the Deeplake dashboard.
OpenClaw config changes
The plugin modifies ~/.openclaw/openclaw.json in two places, both triggered by explicit user commands and both with timestamped backups:
/hivemind_setupappends"hivemind"totools.alsoAllowso OpenClaw admits the plugin's agent tools. OpenClaw's defaultcodingprofile only exposes core tools (read/write/exec/etc.) to agents; plugin-registered tools are filtered out unless explicitly allowed./hivemind_autoupdate [on|off]setsplugins.entries.hivemind.config.autoUpdate. When on, the plugin adds a short line to the system prompt when a newer version is available on ClawHub; the actual install runs through the agent's existingexectool or viaopenclaw plugins update hivemindin a terminal.
The plugin does not replace the built-in memory plugin. It runs alongside memory-core via lifecycle hooks, so memory-core's dreaming cron and other memory-slot jobs keep working.
Troubleshooting
Hivemind feels slow or makes tools hang.
Check agents.defaults.model in ~/.openclaw/openclaw.json. Memory-heavy workflows issue many small tool calls; a large reasoning model feels sluggish. Recommended default is anthropic/claude-haiku-4-5-20251001.
Model switch rejected as "not allowed".
OpenClaw's allowlist wants <provider>/<exact-dated-id>. Use anthropic/claude-haiku-4-5-20251001 or anthropic/claude-sonnet-4-6. Bare IDs and -latest suffixes are rejected.
openclaw model <id> fails with "plugins.allow excludes model".
The CLI is disabled by default. Edit ~/.openclaw/openclaw.json under agents.defaults.model and restart the gateway: systemctl --user restart openclaw-gateway.service.
Telegram-triggered sudo npm i -g openclaw@latest fails with "elevated is not available".
tools.elevated.allowFrom.telegram isn't set. Run the upgrade in a local shell instead.
Sharing memory with teammates
Invite teammates to your Deeplake org:
invite alice@example.com as admin
Their agents will see your memory; your agents will see theirs.