Hivemind shares coding-agent traces across a team

Activeloop has introduced Hivemind, a system that records coding-agent sessions, extracts recurring patterns, and shares the resulting skills across a team. The project is aimed at assistants such as Claude Code, OpenClaw, Codex, Cursor, Hermes Agent, and pi.

According to the project description, Hivemind captures prompts, tool calls, responses, and other session activity as structured traces in Deeplake. It then mines those traces for repeated behavior and converts them into reusable skills that can be made available to other agents in the same workspace.

The company describes the workflow as “capture, codify, propagate, compound.” In practice, that means one agent can encounter a tricky task, and later agents can reuse the pattern without re-solving it from scratch.

Hivemind also includes search across traces and skills. When the local index is unavailable, it falls back to lexical retrieval, with grep-style searching as a backup.

The tool is installed through a single command and can detect supported assistants on the machine. Activeloop says the installer wires up hooks or extensions depending on the platform, opens a browser for login, and then requires the assistants to be restarted.

For Claude Code, Hivemind is available through the native plugin marketplace. For OpenClaw, Activeloop says users can install the ClawHub plugin and then sign in with the /hivemind_login command.

The project also stores data in the same sessions table used by Claude Code and Codex, and the company says it does not take over OpenClaw’s built-in memory-core plugin. That means memory-core keeps handling its own recall and dreaming jobs while Hivemind handles capture and its own commands.

Hivemind’s documentation also describes session summaries, generated by a background worker at the end of each session. Those summaries are written as AI-generated wiki pages and stored in the memory table. Long sessions are checkpointed during use, according to the project notes.

The system stores captured data in Deeplake’s workspace, and the project says all users in that workspace can read it. A data notice appears at the start of each session, and workspace-level isolation is intended to keep orgs separate.

Activeloop says Hivemind can also surface useful questions for teams, such as what someone was working on or what the team decided about an API design. The company positions the tool as a way to make one agent’s work available to the rest of the team through shared traces and shared skills.

The project is open source under the Apache License 2.0, and the source repo includes installers, hook integrations, and documentation for supported assistants. Activeloop also lists planned additions, including trajectory export for fine-tuning, GPU-accelerated retrieval, and skill review before org-wide rollout.