neurohive

Multi-agent memory intelligence — shared knowledge, expertise tracking, and conflict detection for AI agent teams.

Why

AI agents in teams forget everything and don't share knowledge. Multiple agents working on the same project re-discover the same facts, contradict each other, and have no collective memory. One agent learns the API timeout is 30 seconds; another assumes 60. Both conclusions sit in separate sessions, never compared.

neurohive wraps neuromcp with a shared intelligence layer. Every memory stored by any agent becomes visible to the team. Searches come back annotated with who discovered what and when. New agents onboard with the team's existing knowledge already loaded. Contradictions get flagged before they cause incidents.

Before & After

Quick Start

npx neurohive

Drop-in replacement for npx neuromcp. Use the same MCP config — just swap the command. All neuromcp tools and resources are available unchanged, plus three new neurohive tools, five new resources, and automatic cross-agent annotation.

Features

Cross-Agent Annotations

Every search result is annotated with provenance: which agent stored it, how long ago, and how many times it's been retrieved. Agents can see at a glance whether a memory came from a trusted teammate or an unknown source.

Agent Priming

When an agent starts up, it reads the hivemind://priming resource to load the team's top knowledge, recent activity, and known gaps. No cold starts.

Expertise Tracking

Every time an agent stores a memory, neurohive updates that agent's expertise profile — tracking which categories they contribute to most. Use hivemind_expertise to route questions: "who knows the most about authentication?"

Conflict Detection

When two agents store contradictory claims about the same fact, neurohive flags it as a conflict. Conflicts surface in hivemind_conflicts and in the dashboard so the team can resolve them explicitly rather than letting stale data propagate silently.

Dashboard

A single command gives a live view of the collective knowledge base: total memories, active agents, top categories, recent activity, unresolved conflicts, and knowledge gaps.

Installation

Claude Code

{
  "mcpServers": {
    "neurohive": {
      "command": "npx",
      "args": ["neurohive"],
      "env": {
        "HIVEMIND_COMPANY": "my-project",
        "HIVEMIND_AGENT": "Agent-Backend",
        "HIVEMIND_AGENT_ROLE": "backend"
      }
    }
  }
}

Claude Desktop

{
  "mcpServers": {
    "neurohive": {
      "command": "npx",
      "args": ["-y", "neurohive"],
      "env": {
        "HIVEMIND_COMPANY": "my-project",
        "HIVEMIND_AGENT": "Agent-Frontend"
      }
    }
  }
}

Cursor

{
  "mcpServers": {
    "neurohive": {
      "command": "npx",
      "args": ["neurohive"],
      "env": {
        "HIVEMIND_COMPANY": "my-project",
        "HIVEMIND_AGENT": "Cursor-Agent"
      }
    }
  }
}

Paperclip Mode

When Paperclip is running locally, neurohive detects it automatically and reads the company, agent, and role from the Paperclip API. No extra env vars needed — the company namespace is set from the active Paperclip company, and each agent is identified by its Paperclip identity.

Manual fallback (when Paperclip is not running):

HIVEMIND_PAPERCLIP_URL=http://localhost:3100   # default
HIVEMIND_COMPANY=my-project
HIVEMIND_AGENT=Agent-Backend
HIVEMIND_AGENT_ROLE=backend

CLI Commands

Configuration

neurohive variables

neuromcp pass-through

All NEUROMCP_* environment variables pass through unchanged. Key ones:

MCP Surface

neurohive exposes 16 tools, 18 resources, and 3 prompts.

Tools

13 neuromcp tools (unchanged):

3 neurohive tools:

Resources

13 neuromcp resources (memory, graph, priming, stats) plus 5 neurohive resources:

Prompts

3 neuromcp prompts: store_memory_prompt, search_memory_prompt, session_summary_prompt.

Paperclip Integration

neurohive probes HIVEMIND_PAPERCLIP_URL on startup. If the Paperclip API responds:

1. Namespace is set to company-<companyId> — all agents on the same Paperclip company share one knowledge base.

2. Agent identity is read from the Paperclip agent record, so each agent in a fleet is automatically identified.

3. Role is pulled from the agent definition, used to enrich priming context ("you are the backend agent; here's what your team knows").

If Paperclip is not running, neurohive falls back to HIVEMIND_COMPANY / HIVEMIND_AGENT env vars with no loss of functionality.

Requirements

Platform Support

FAQ

Is this a replacement for neuromcp? Yes and no. neurohive runs on top of neuromcp and re-exports all its tools, resources, and prompts. If you run a single agent, it's functionally identical to neuromcp with a few extra tools. The value emerges with two or more agents sharing the same HIVEMIND_COMPANY namespace.

Do all agents share one database? Yes — agents in the same company namespace share a database via the shared filesystem path (NEUROMCP_DB_PATH). This works naturally in multi-agent setups on the same machine (Paperclip, dmux, parallel Claude Code sessions). For distributed setups, point all agents to a shared NFS or networked path.

What counts as a conflict? When two memories from different agents make contradictory claims about the same subject. Conflict detection runs at store time using the cognitive claims extracted from memory content. You can tune sensitivity or disable it via the neuromcp config.

Will it break my existing neuromcp setup? No. neurohive passes all existing NEUROMCP_* env vars through unchanged and writes to the same database. Switching back to npx neuromcp works at any time.

How does priming work? At startup, agents that read hivemind://priming receive a formatted summary: team size, top categories, most active agents, recent memories, and any open conflicts. This is designed to be included in a system prompt or read via a hook on session start.

Contributing

See CONTRIBUTING.md.

License

MIT — see LICENSE.