Hivemind — Shared Brain for Your Agents
Turn your team's agent traces into reusable skills that propagate across every agent.
Evidence: one-line consent prompt during install, data collection notice clearly lists captured data types, and options to disable capture (HIVEMIND_CAPTURE=false, .hivemind collect:false). Credentials file stored with 0600 permissions, directory 0700. However, least privilege is insufficient: installer modifies multiple agent config files, and capture is enabled by default without fine-grained permission control. User confirmation: one-line consent prompt during install, but no explicit confirmation of data collection scope. Data flow transparency: README details captured data types and storage location, but not how data is transmitted to cloud. Sensitive data handling: credentials file permissions correct, but no mention of encryption or anonymization. Dependency security: dependency versions not pinned, potential known vulnerabilities. External effects: installer modifies multiple agent configs, but no rollback mechanism described. Rollback: uninstall command provided, but impact on configs not explained. Source attribution: no clear source attribution information.
Self-consistency: README consistent with code structure, features match implementation. Dependency availability: dependency versions not pinned, availability risk. Failure messages: tests cover error handling, but no user-visible error message examples.
Audience and scenarios: clearly supports multiple agents, provides install guides for different scenarios. Capability boundaries: clearly states capture scope (e.g., Claude Cowork only Local Agent Mode), but other limitations not mentioned. Trigger precision: capture controlled via env vars and config files, but trigger conditions not specified. Environment fit: supports multiple OS and agents, but no specific environment limitations.
Information architecture: README well-structured, but missing API docs. Install notes: multiple install methods, but no troubleshooting. Naming stability: version numbers update frequently, but naming conventions not explained. Examples and FAQ: examples provided, but no FAQ. Known limitations: Claude Cowork limitation clearly stated, but others not. License: Apache-2.0, but no copyright info. Versioning and changelog: no changelog. Maintenance responsibility: maintainers not identified.
Output usability: skills and search provided, but output format not specified. Marginal value: shared memory and skill reuse, but benefits not quantified. Cost-benefit: benchmark data provided, but actual costs not explained.
Claim traceability: benchmark data in README has sources, but no reproduction method. Cross-source corroboration: only README, no other sources. Fact-inference separation: README distinguishes facts and inferences, but not explicitly labeled.
- Capture is enabled by default, may inadvertently collect sensitive data; users should configure explicitly.
- Dependency versions are not pinned, posing supply chain risk.
- Installer modifies multiple agent configs; evaluate permissions carefully.
What does this agent do, and when should you use it?
Hivemind is a cloud-backed shared memory engine for coding agents like Claude Code, OpenClaw, Codex, Cursor, Hermes, and pi. It automatically captures prompts, tool calls, and responses as structured traces in Deeplake, then mines them for recurring patterns to codify into reusable SKILL.md files. Features include hybrid lexical+semantic retrieval, proactive recall, session summaries, a codebase graph, and automatic rule sharing. Data is stored in your Deeplake workspace by default, with BYOC support for GCS, Azure, or S3. Privacy controls via .hivemind files allow per-directory routing or opt-out. Licensed under Apache-2.0.
Hivemind installs hooks or extensions into supported agents (Claude Code, OpenClaw, Codex, Cursor, Hermes, pi) to capture every session's prompts, tool calls, and responses, writing them to the Deeplake sessions table. A background worker generates AI-written wiki summaries (stored in the memory table) and mines for patterns to create SKILL.md files in the project's .claude/skills/ directory. During inference, hybrid retrieval (BM25 + optional embeddings) and proactive recall inject relevant memories into the agent's context. It also virtualizes ~/.deeplake/memory/ as a file system backed by SQL, builds a codebase graph of files/symbols/imports, and can generate code docs and team rules using the host agent's CLI (e.g., claude -p).
- A team wants new engineers to benefit from past solutions discovered by senior engineers' agents.
- A Claude Code user asks 'What was Emanuele working on?' and gets answers from shared memory.
- OpenClaw users want session capture and skill mining without conflicting with the built-in memory-core plugin.
- Codex users need cross-session memory without manually managing hooks.json.
- An org wants to enforce rules like 'no DROP TABLE on prod' across all coding agents.
- A consultancy needs to isolate client work by routing traces to specific workspaces via .hivemind files.
What are this agent's strengths and limitations?
- Measured improvements on LoCoMo benchmark: 25% cheaper, 1.7× fewer tokens, 31% fewer turns
- Multi-agent support (Claude Code, OpenClaw, Codex, Cursor, Hermes, pi) with automated hook installation
- Automatic skill mining and propagation across agents
- Hybrid retrieval (BM25 + embeddings) and proactive recall
- BYOC support for GCS, Azure, S3, and on-prem storage
- Granular privacy controls via .hivemind routing and opt-out
- Default reliance on Deeplake cloud (API, account) – not fully self-hosted
- All workspace members can read captured data and skills – requires trust
- Semantic search requires an additional ~600 MB embedding model, otherwise degrades to lexical only
- OpenClaw integration may feel slow with large reasoning models; recommended to use Haiku
- Codex requires manual hook trust on first launch; otherwise inactive
- Claude Cowork integration is Alpha and only captures Local Agent Mode sessions, not desktop chat
How do you install or deploy this agent?
Install globally: npm i -g @deeplake/hivemind && hivemind install. This detects installed agents and wires hooks, showing a one-line consent prompt before signing in via browser. For headless installs: HIVEMIND_TOKEN=<token> hivemind install or hivemind install --token <token>. You can install for a specific agent: hivemind claude install, hivemind codex install, hivemind claw install, etc. Restart your agents after install. Requires Node.js >=22 and a Deeplake account (get a token at https://deeplake.ai).
How do you use this agent?
Once installed, agents automatically capture sessions. Check status with hivemind status. Query memory naturally, e.g., 'Search traces for authentication bugs we've solved'. Manage skills: hivemind skillify shows current scope, hivemind skillify scope <me|team> sets mining scope, hivemind skillify pull installs teammates' skills. View summaries in ~/.deeplake/memory/summaries/. Disable capture with HIVEMIND_CAPTURE=false claude. Configure per-directory routing by placing a .hivemind file like { "orgId": "acme-corp", "workspaceId": "client-work" }. Enable semantic search with hivemind embeddings install (adds ~600 MB dependency).
How does this agent compare with similar options?
Hivemind coexists with OpenClaw's built-in memory-core plugin without claiming the memory slot. No direct comparisons to other memory engines like MemGPT or Mem0 are mentioned.
FAQ
Is Hivemind free?
How secure is my data?
Can I control which sessions are captured?
HIVEMIND_CAPTURE=false to disable capture entirely, or place a .hivemind file with collect: false in a directory to opt out. Use .hivemind.local for personal overrides.