One brain for all your agents, Hivemind provides a shared, auto-learning knowledge base for AI agents, enabling teams to leverage collective intelligence. It captures agent interactions, codifies reusable patterns into skills, and propagates this knowledge across all team agents in real-time, significantly reducing redundant work, cutting costs, and improving agent efficiency.

How It Works

The system captures every prompt, tool call, and response as structured traces in Deeplake. A background worker mines these traces for recurring patterns, codifying them into SKILL.md files. These skills are then automatically propagated to every Hivemind-connected agent at inference time, enhancing their capabilities based on team-wide learnings. It employs hybrid lexical and semantic retrieval for efficient searching of traces and skills.

Quick Start & Requirements

• Primary Install: npm install -g @deeplake/hivemind && hivemind install
• Prerequisites: Requires supported AI agents such as Claude Code, OpenClaw, Codex, Cursor, Hermes Agent, or pi. For optimal performance with OpenClaw, anthropic/claude-haiku-5-20251001 is recommended. Optional GPU acceleration for dense retrieval is available.
• Links:
  • API Token: https://deeplake.ai
  • Docs: docs/ARCHITECTURE.md, docs/EMBEDDINGS.md, docs/SUMMARIES.md, docs/SKILLIFY.md

Highlighted Details

• Benchmarks: On the LoCoMo long-context memory benchmark, Hivemind achieves 25% lower cost, 1.7x fewer tokens, and 31% fewer turns compared to a baseline without shared memory.
• Features: Real-time propagation of codified skills, hybrid semantic/lexical search, AI-generated session summaries, and a virtual filesystem for memory management.
• Codebase Graph: Constructs a live graph of the codebase from agent interactions, enabling recall based on traversed file paths and symbols, not just text mentions.
• Team Rules: Facilitates sharing of cross-agent team principles, injected into agent sessions to enforce team policies.

Maintenance & Community

Hivemind is developed and maintained by Activeloop, the team behind Deeplake, with backing from Y Combinator. For inquiries or support, contact hello@activeloop.ai or join their community channels. The roadmap includes GPU-accelerated retrieval and enhanced skill versioning.

Licensing & Compatibility

The project is licensed under the Apache License 2.0, permitting commercial use and integration into closed-source projects without significant copyleft restrictions.

Limitations & Caveats

The optional local embedding daemon for semantic search requires ~600MB of disk space and is disabled by default; search degrades to lexical-only without it. Data capture is comprehensive, including all prompts, tool calls, and responses, stored in a shared Deeplake workspace accessible to all users within that workspace. For optimal responsiveness, specific, smaller agent models are recommended.