activeloopai/hivemind

Hivemind turns your traces into reusable skills across agents

What it solves

Hivemind provides a shared, cloud-backed memory system for AI coding agents. It prevents the "forgetting" problem where every agent session starts from scratch, and ensures that when one agent (or teammate) solves a complex problem, the solution is automatically available to all other agents on the team.

How it works

Hivemind operates through a cycle of capture, codification, and propagation:

  1. Capture: It intercepts prompts, tool calls, and responses from supported agents as structured traces stored in Deeplake.
  2. Codify: Background workers analyze these traces to identify repeated patterns and extract them into reusable SKILL.md files.
  3. Propagate: These codified skills and session summaries are made available to all connected agents via hybrid lexical and semantic search, allowing agents to recall prior work during inference.
  4. Virtual Filesystem: It uses a virtual filesystem backed by SQL to intercept file operations on ~/.deeplake/memory/.

Who it’s for

Developers and engineering teams using AI coding assistants such as Claude Code, Cursor, OpenClaw, Codex, Hermes, or pi who want to share knowledge and reduce token spend and session turns across their team.

Highlights

  • Cross-Agent Memory: Works across different AI assistants on the same team, allowing knowledge to propagate in real time.
  • Pattern Mining: Automatically turns successful agent interactions into codified skills (SKILL.md).
  • Cost Efficiency: Demonstrated reduction in costs (25%), token usage (1.7x fewer), and turns per question (31% fewer) on the LoCoMo benchmark.
  • Broad Integration: Supports a wide range of agents via plugins, hooks, and MCP servers (including an Alpha integration for Claude Cowork).
  • BYOC Storage: Allows users to keep their data in their own GCS, Azure, S3, or on-prem buckets.

Related

  • Project
    24kchengYe/MemoMind

    MemoMind is a locally hosted, GPU‑accelerated memory system for AI coding agents. It stores extracted facts from chats, documents, and daily life events in PostgreSQL + pgvector, builds a knowledge graph, and offers fast 4‑way hybrid retrieval (semantic, BM25, graph, temporal). The agent can retain new information, recall relevant memories, and reflect across the whole store. All data stays on the user’s machine, works with any OpenAI‑compatible LLM, and includes a web dashboard for browsing and exporting memories.

  • Project
    zilliztech/memsearch

    memsearch is a Python library and CLI that records AI coding‑assistant chats as Markdown, builds a hybrid (BM25 + dense) Milvus index, and provides cross‑platform recall via plugins for Claude Code, Codex, DeepSeek Harness, OpenClaw, and OpenCode. It supports local ONNX embeddings, optional Zilliz Cloud backend, background project/user notes, and automatic skill distillation.

  • Project
    AMAP-ML/SkillClaw

    SkillClaw is an open‑source system that sits between LLM agents and the LLM API, automatically records interactions, extracts reusable “skills”, deduplicates them, and continuously refines them. It works with many agents (Hermes, Codex, Claude Code, QwenPaw, OpenClaw, etc.) and can share a unified skill library across devices, agents, and team members via local storage, OSS, or S3.

  • Project
    supermemoryai/supermemory

    Supermemory is a hosted‑plus‑self‑hostable memory and context engine for LLM‑based assistants. It extracts facts from chats, emails, PDFs, images, code, etc., builds per‑user profiles (static facts + dynamic recent context), and provides hybrid RAG + memory search through a single API (`add`, `profile`, `search`). Plugins let popular tools (Claude Code, Cursor, Codex) gain persistent memory, while a one‑binary `supermemory local` server lets you run the whole stack offline with local embeddings. The project claims state‑of‑the‑art performance on LongMemEval, LoCoMo and ConvoMem benchmarks (95 % Recall@15 with >99 % context reduction).

  • Dispatch
    funes: durable memory layer for coding agents

    Hugging Face announced funes, a single‑binary, locally‑run memory layer that indexes coding‑agent session traces and lets agents recall raw evidence across runs, machines, and models.