# Hivemind

An auto-learning, cloud-backed shared memory system that mines coding-agent traces for repeated patterns and turns them into skills usable by every agent on a team.

Hivemind is a shared memory layer for coding agents such as Claude Code, OpenClaw, Codex, Cursor and Hermes, built so a pattern one engineer's agent discovers becomes available to every other agent on the team soon afterward. It captures each session's prompts, tool calls and responses as structured traces stored in Deeplake, then mines those traces for repeated patterns and codifies them into reusable SKILL.md files that propagate across sessions, agents, teammates and machines in real time.

Traces and skills are searched with hybrid lexical and semantic retrieval, with a lexical fallback when embeddings are turned off, and a background worker summarizes sessions into AI-generated wiki pages. On the LoCoMo long-context memory benchmark, the project reports meaningful reductions in cost, tokens and turns compared with running the same agent without shared memory. Data can be kept in the user's own cloud storage bucket or on-premises, and it installs with a one-line script for macOS, Linux, Windows or npm-based CI setups. It is written in TypeScript.

Website: https://deeplake.ai/hivemind