Correct one agent.Every agent learns.

Hivemind is persistent memory across your team's AI agents. Tell one agent it got something wrong; once your lead approves the lesson, every agent follows it.

Watch it work Open the app

In our eval, agents broke the team's rule in 22 of 36 runs without Hivemind, and in 0 with it. See the results

01@kevin corrects Claude Code

We built Hivemind for AI training-data teams like AfterQuery's. The company in the demo, Meridian Data, is fictional.

Bad lessons never reach an agent.

Security locks the rules that matter. A lesson that contradicts one is blocked, and nobody below Security can change it.

Blocked by a locked rule· @emily's agent, just now

@emily tried to teach

“Just reuse Lab A's rows for the Lab B test fixtures, it's only tests”

vs

Locked by @greg, Head of Security

Never mix data from different lab customers

You're the lead. Approve it.

Kevin's correction is waiting in your queue. Press a (or tap Approve), the same key as in hivemind watch.

ben — hivemind watch

⬢ Hivemind · Meridian Data / export-service      ● @ben  Lead   
 Here   Rules 10   Queue 2   Feed                               
──────────────────━━━━━━━━━─────────────────────────────────────
Waiting on you · 2                                              
▎ Every export calls assertSingleCustomer()                     
  "No. Exports must call assertSingleCustomer() before writing  
  anything."                                                    
   ● #team/data-pipelines   2m                                  
  proposed by  ● @kevin's  Claude Code agent                    
  a approve   r reject                                          
▎ Retries back off exponentially                                
   ● #repo/export-service   18m                                 
  proposed by  ● @emily's  Cursor agent                         
STOPPED BY A LOCKED RULE · 1                                    
▎ Mixing labs is fine for test fixtures                         
  Blocked: contradicts  ● @greg's  locked rule                  
────────────────────────────────────────────────────────────────
a approve   r reject   j/k move   enter open                    

Your team's agents

team rules 10

@kevinClaude Codeknows it
@emilyCursorknows it
@tomCodexknows it
@laurenGemini CLIknows it
@benCopilotknows it

Waiting on you

Exports call assertSingleCustomer()

or press a

What happens while you work.

One afternoon on the export service, in order. All of it in the terminal, next to whichever agent you use.

2:04 pmKevin's agent saves src/payouts.ts
It catches a bad save in under a second
hivemind watch, in the pane next to the agent, underlines the exact code that breaks a rule and offers the fix. hivemind check --staged does the same before a commit.
BREAKS A LOCKED RULE · #org
Never log expert personal info
2 │ console.log("sent", expert.email);
│ ~~~~~~~~~~~~ personal info
f fix: log the expert id instead
2:06 pmKevin fixes it by hand
His fix becomes a lesson
Hivemind reads the diff, writes the lesson and sends it to Ben, his lead. Once Ben approves, every agent on the team follows it.
- console.log(user.email);
Hivemind · Lesson captured from @kevin's fix in src/welcome.ts:2: "Log the expert id, not their email". Waiting on @ben.
3:40 pmEmily asks her agent for something risky
The agent says no, and she asks why
It names the locked rule and offers a safe way to do it. /hivemind why lays out the rule's history: who set it, the incident behind it, and what it has blocked.
> Log the expert's email when a payout fails so I can debug it.
● Blocked by Hivemind · LOCKED RULE "Never log expert personal info" · @greg
Expert emails live only in the pii schema, so they can't be written to the payout handler's logs.
Instead: log the hashed expert id with the trace fields … logger.info("payout failed", { trace_id, tenant_id, job_id, expert: hmacExpertId(expertId) })
Run /hivemind why for the history.
> /hivemind why
⎿ ⬢ Hivemind · why
LOCKED RULE · #org
Never log expert personal info
Set by @greg · Head of Security · 16 Sep · Incident INC-27
"A stack trace with an expert's email reached our error tracker. Never again."
In context every agent in #repo/export-service at session start · seen today: Claude Code
Reinforced 23× · 1 evidence · always applies · blocked 1 proposal
╰ latest: "Log the expert's email when a payout fails" from @emily 1h ago
Only @greg can change this rule · open in Hivemind
4:15 pmTwo agents reach the same file
It warns before the second one edits
Whatever tool the other person uses, your agent hears who changed the file in the last 10 minutes.
● Update(src/export.ts)
Hivemind · Heads up: @kevin's Cursor agent edited src/export.ts 2 min ago.
05
It works with the agents you already use
Claude CodeCursorCodexGitHub CopilotGemini CLIWindsurf
hivemind init connects a repo to all of them. !hivemind opens watch next to Claude Code.

Lessons live where they apply.

Company rules reach everyone, team lessons reach the team, and your own stay yours. Zoom from the company to one lesson.

An agent in #repo/export-service starts every session with 9 rules(locked rules first, then from the company down to here)
Never mix data from different lab customers#org
Never log expert personal info#org
Money and payouts are integer cents#org
Use logger, not console.log#org
Migrations are append-only#org
Tag every row with lab_customer_id#data-pipelines
Ask @ben about export-service#data-pipelines
Export jobs are idempotent, keyed by job_id#export-service
Stream uploads; never buffer a full export#export-service

Set up in one command.

hivemind init finds the agents a repo already uses and connects each one. Then type !hivemind in Claude Code to open watch next to it.

~/export-service — zsh

~/export-service $ hivemind init
⬢ hivemind init
AgentDetected viaMCP / hooks configRules file
✓ Claude Code.claudecreated .mcp.jsonSessionStart hook
✓ Cursor.cursorcreated .cursor/mcp.json.cursor/rules/hivemind.mdc
✓ Gemini CLIGEMINI.mdcreated .gemini/settings.jsonGEMINI.md
✓ Codex CLIAGENTS.mdcreated .codex/config.tomlAGENTS.md
✓ GitHub Copilot.vscodecreated .vscode/mcp.json.github/copilot-instructions.md
✓ Windsurf.windsurfcreated .windsurf/mcp_config.hivemind.example.json.windsurf/rules/hivemind.md
· created .hivemind.json

Agent	Detected via	MCP / hooks config	Rules file
✓ Claude Code	.claude	created .mcp.json	SessionStart hook
✓ Cursor	.cursor	created .cursor/mcp.json	.cursor/rules/hivemind.mdc
✓ Gemini CLI	GEMINI.md	created .gemini/settings.json	GEMINI.md
✓ Codex CLI	AGENTS.md	created .codex/config.toml	AGENTS.md
✓ GitHub Copilot	.vscode	created .vscode/mcp.json	.github/copilot-instructions.md
✓ Windsurf	.windsurf	created .windsurf/mcp_config.hivemind.example.json	.windsurf/rules/hivemind.md

Before the agent writes, not after review.

Other tools catch a mistake after it's written, or only for one vendor's agent.

Hivemindreview bots

Session starts
Agent writes
File saved
PR opened
Code review

Review bots Greptile, CodeRabbitact after a PR is open
Vendor memory Copilot Memory, Cursor ruleshelp only their own agent
Rule files CLAUDE.md, AGENTS.mdhold only what someone wrote down
Memory libraries Mem0, Zep, Lettaare building blocks, with no approvals
Hivemindacts before the agent writes, for every agent, approved by your lead

Correct it once.

Every agent on your team remembers.

Open the app See how it works

Correct one agent.Every agent learns.

One correction reaches every agent.

One correction reaches every agent.

Kevin corrects his agent

Ben approves it with one key

Every agent has it

Emily's agent gets it right

Bad lessons get blocked

Bad lessons never reach an agent.

You're the lead. Approve it.

What happens while you work.

It catches a bad save in under a second

His fix becomes a lesson

The agent says no, and she asks why

It warns before the second one edits

It works with the agents you already use

Lessons live where they apply.

Set up in one command.

Before the agent writes, not after review.

Correct it once.