Skip to main content

4.2 HiveMind — Shared Knowledge

HiveMind — Shared Knowledge Base

HiveMind (~/system/agents/hivemind/hivemind.js) is the collective memory of all agents.

What Goes Into HiveMind

  • Task completions — every pi-orchestrator task result (Tier 2+)
  • Agent discoveries — facts, patterns, warnings found during work
  • Session summaries — condensed logs of work sessions
  • Manual knowledge — docs, runbooks, architecture decisions

How It Works

Agent completes task
       │
       ▼
HiveMind stores entry:
  - Text (up to 5000 chars)
  - Source (which agent/engine)
  - Type (knowledge, alert, briefing, update)
  - Timestamp
       │
       ▼
BGE-M3 embeds the entry → Qdrant vector DB
       │
       ▼
Future agents query HiveMind:
  "What do we know about Drop's payment system?"
       │
       ▼
Semantic search → relevant entries returned as context

Usage

# Query knowledge
node hivemind.js query "How does Drop handle PSD2 consent?"

# Store new knowledge
node hivemind.js post my-agent knowledge "Drop uses SCA for all transactions over 30 EUR"

# Search by tag
node hivemind.js search --type alert --since 24h

Why It Matters

Without HiveMind, every agent starts from zero. With HiveMind:

  • Agents learn from previous work
  • Duplicate investigation is avoided
  • Institutional knowledge persists across sessions
  • Quality improves over time (flywheel effect)
Back to top