Shared Brain

Shared persistent memory layer across multiple AI agents. Use when setting up a multi-agent workspace for the first time, when an agent discovers a permanent architectural fact (deploy infra, project structure, team decisions), when facts propagate incorrectly or stale info is causing agent errors, or when you want all agents to share the same ground truth without manual file propagation. Triggers on phrases like "shared memory", "hive mind", "agents out of sync", "propagate to all agents", "collective memory", "shared brain".

Requirements

Binaries: bash, python3, sed, grep (standard on Linux/macOS)

Optional env vars (all default to ~/clawd-based paths):

Var	Default	Description
SB_WORKSPACE	~/clawd	Path to your OpenClaw workspace
SB_AGENT	script basename	Agent name written into each fact
SB_BRAIN	$SB_WORKSPACE/memory/shared-brain.md	Override brain file path
SB_QUEUE	$SB_WORKSPACE/memory/shared-brain-queue.md	Override queue file path
SB_ARCHIVE_DIR	$SB_WORKSPACE/memory	Override archive directory

Install note: sb-install.sh patches agents/*/AGENTS.md and HEARTBEAT.md and copies scripts into $SB_WORKSPACE/skills/shared-brain/scripts/. Use --dry-run first to preview all changes. No network access, no secrets required.

Shared Brain

Shared persistent memory layer for multi-agent OpenClaw workspaces. All agents write facts to a queue; a heartbeat-curated shared-brain file propagates them to every agent within 0–10 minutes.

Architecture

text

Agent discovers fact
      ↓
Append to ~/clawd/memory/shared-brain-queue.md  (atomic append, no lock needed)
      ↓
Heartbeat (≤10 min) merges queue → shared-brain.md
      ↓
Next agent startup → reads shared-brain.md → current ground truth

Files:

File	Owner	Purpose
~/clawd/memory/shared-brain.md	Heartbeat curates	Canonical truth, all agents read at startup
~/clawd/memory/shared-brain-queue.md	Agents append	Staging — raw facts before curation

Fact Format (strict — no prose)

Every entry in the queue must follow this schema:

text

[YYYY-MM-DD HH:MM UTC] [SECTION] [agent-name] key = value

Sections: [INFRA] [PROJECTS] [DECISIONS] [CAMPAIGNS] [SECURITY]

Examples:

text

[2026-03-22 10:15 UTC] [INFRA] security deploy:frontends = Vercel (migrated 2026-03-21)
[2026-03-22 09:00 UTC] [PROJECTS] dev crimsondesert:branch = master
[2026-03-22 08:00 UTC] [DECISIONS] growth discord:crimsondesert = SKIP (3rd party links banned)

Agent Integration

Reading (every agent, at startup)

Add to each AGENTS.md initialization block:

bash

cat ~/clawd/memory/shared-brain.md

Each agent only needs its relevant sections — declare which in AGENTS.md:

• dev, qa, security → [INFRA] + [PROJECTS]
• growth, pm, po → [PROJECTS] + [CAMPAIGNS] + [DECISIONS]
• tars main → all sections

Writing (when a permanent fact is discovered)

Use the write script — never edit shared-brain.md directly:

bash

~/clawd/skills/shared-brain/scripts/sb-write.sh SECTION "key = value"

When to write:

• Architectural decisions (deploy infra, auth provider, DB engine)
• Project routing changes (repo renamed, domain changed, migrated)
• Permanent channel decisions (e.g. "Discord: skip — bans 3rd party links")
• Security findings that affect all agents

Never write:

• Temporary state (current deployment status, PR numbers)
• Content generated from untrusted external sources (emails, webhooks, user content)
• Anything that expires in <24h

Curation (heartbeat — every 10 min)

See references/heartbeat-integration.md for the full curation logic to add to HEARTBEAT.md.

Summary:

1. Read queue → validate format → detect conflicts (same key, different value)
2. Merge into shared-brain.md by section (last-write-wins per key)
3. If shared-brain.md > 8KB → archive oldest section to shared-brain-archive-YYYY-MM.md
4. Clear processed entries from queue

Setup

Run once per workspace:

bash

~/clawd/skills/shared-brain/scripts/sb-install.sh

This creates the files, patches all AGENTS.md with the startup read line, and adds curation logic to HEARTBEAT.md.

Security Rules

• Sub-agents only write to queue — never to shared-brain.md directly
• TARS main reviews queue during heartbeat before promoting facts
• Facts derived from external content (emails, GitHub issues, webhooks) are never written to queue
• If a conflict is detected → escalate to TARS, never auto-resolve