Hivemind

Shared memory for a team's AI coding agents. When an engineer corrects their agent, Hivemind turns the correction into a short lesson. A lead approves it, and every teammate's agent (Claude Code, Cursor, Codex, Copilot, Gemini CLI) loads it at the start of its next session. Security can lock rules so a lesson that contradicts them never reaches an agent.

I built the learning engine and the agent integrations: Claude Haiku decides whether a correction is a reusable rule, merges it with existing lessons and checks it against locked rules. Lessons are ranked by a confidence score that goes up when they hold and down when they get corrected.

In an eval of 12 coding tasks with 3 runs each, graded by code, headless Claude Code broke the team's unwritten rules in 22 of 36 runs with no memory and 0 of 36 with Hivemind's lessons.

Live site ↗︎ GitHub ↗︎