brainx（brainx 搜索）

BrainX V5 — The First Brain for OpenClaw

Persistent memory system using vector embeddings for contextual retrieval in AI agents.

37 Features

When to Use

✅ USE when:

• An agent needs to "remember" information from previous sessions
• You want to give additional context to an LLM about past actions
• You need semantic search by content
• You want to store important decisions with metadata

❌ DON'T USE when:

• Ephemeral information that doesn't need persistence
• Structured tabular data (use a regular DB)
• Simple cache (use Redis or in-memory)

Auto-Injection (Hook)

BrainX V5 includes an OpenClaw hook that automatically injects relevant memories when an agent starts.

Production Validation Status

Real validation completed on 2026-03-18:

• Global hook enabled in ~/.openclaw/openclaw.json
• Managed hook synced with ~/.openclaw/skills/brainx-v5/hook/ (handler.js re-synced)
• Active physical database: brainx_v5
• agent-profiles.json expanded from 10 to 32 profiles (all agents)
• Cross-agent injection slots (30%) activated in production
• 20 null embeddings regenerated + 17 duplicate pairs deduped via brainx fix
• 2 pending migrations applied
• Doctor: 18/18 passed, 0 warnings
• Real bootstrap smoke test passed for 10 agents
• Expected evidence confirmed:

- <!-- BRAINX:START --> block written into MEMORY.md - Updated: timestamp present - Fresh row recorded in brainx_pilot_log

If this validation becomes stale, rerun a bootstrap smoke test before assuming runtime is still healthy.

How it works:

1. agent:bootstrap event → Hook fires automatically
2. PostgreSQL query → Fetches hot/warm recent memories
3. Generates file → Creates BRAINX_CONTEXT.md in the workspace
4. Agent reads → File is loaded as initial context

Configuration:

In ~/.openclaw/openclaw.json:

{
  "hooks": {
    "internal": {
      "enabled": true,
      "entries": {
        "brainx-auto-inject": {
          "enabled": true,
          "limit": 5,
          "tier": "hot+warm",
          "minImportance": 5
        }
      }
    }
  }
}

Per-agent setup:

Add to AGENTS.md in each workspace:

## Every Session

1. Read `SOUL.md`
2. Read `USER.md`
3. Read `brainx.md`
4. Read `BRAINX_CONTEXT.md` ← Auto-injected context

Available Tools

brainx_add_memory

Saves a memory to the vector brain.

Parameters:

• content (required) — Memory text
• type (optional) — Type: note, decision, action, learning (default: note)
• context (optional) — Namespace/scope
• tier (optional) — Priority: hot, warm, cold, archive (default: warm)
• importance (optional) — Importance 1-10 (default: 5)
• tags (optional) — Comma-separated tags
• agent (optional) — Name of the agent creating the memory

Example:

brainx add --type decision --content "Use embeddings 3-small to reduce costs" --tier hot --importance 9 --tags config,openai

brainx_search

Searches memories by semantic similarity.

Parameters:

• query (required) — Search text
• limit (optional) — Number of results (default: 10)
• minSimilarity (optional) — Threshold 0-1 (default: 0.3)
• minImportance (optional) — Filter by importance 0-10
• tier (optional) — Filter by tier
• context (optional) — Exact context filter

Example:

brainx search --query "API configuration" --limit 5 --minSimilarity 0.5

Returns: JSON with results.

brainx_inject

Gets memories formatted for direct injection into LLM prompts.

Parameters:

• query (required) — Search text
• limit (optional) — Number of results (default: 10)
• minImportance (optional) — Filter by importance
• tier (optional) — Tier filter (default: hot+warm)
• context (optional) — Context filter
• maxCharsPerItem (optional) — Truncate content (default: 2000)

Example:

brainx inject --query "what decisions were made about openai" --limit 3

Returns: Formatted text ready for injection:

[sim:0.82 imp:9 tier:hot type:decision agent:coder ctx:openclaw]
Use embeddings 3-small to reduce costs...

---

[sim:0.71 imp:8 tier:hot type:decision agent:support ctx:brainx]
Create SKILL.md for OpenClaw integration...

brainx_health

Verifies BrainX is operational.

Parameters: none

Example:

brainx health

Returns: PostgreSQL + pgvector connection status.

Backup and Recovery

Create Backup

./scripts/backup-brainx.sh ~/backups

Creates brainx-v5_backup_YYYYMMDD_HHMMSS.tar.gz containing:

• Full PostgreSQL database (SQL dump)
• OpenClaw configuration (hooks, .env)
• Skill files
• Workspace documentation

Restore Backup

./scripts/restore-brainx.sh backup.tar.gz --force

Fully restores BrainX V5 including:

• All memories (with embeddings)
• Hook configuration
• Environment variables

Full Documentation

See RESILIENCE.md for:

• Complete disaster scenarios
• Migration to new VPS
• Troubleshooting
• Automatic backup configuration

Configuration

Environment Variables

# Required
DATABASE_URL=postgresql://user:pass@host:5432/brainx_v5
OPENAI_API_KEY=sk-...

# Optional
OPENAI_EMBEDDING_MODEL=text-embedding-3-small
OPENAI_EMBEDDING_DIMENSIONS=1536
BRAINX_INJECT_DEFAULT_TIER=hot+warm
BRAINX_INJECT_MAX_CHARS_PER_ITEM=2000
BRAINX_INJECT_MAX_LINES_PER_ITEM=80

Database Setup

# Schema is in ~/.openclaw/skills/brainx-v5/sql/
# Requires PostgreSQL with pgvector extension

psql $DATABASE_URL -f ~/.openclaw/skills/brainx-v5/sql/v3-schema.sql

Direct Integration

You can also use the unified wrapper that reads the API key from OpenClaw:

cd ~/.openclaw/skills/brainx-v5
./brainx add --type note --content "test"
./brainx search --query "test"
./brainx inject --query "test"
./brainx health

Compatibility: ./brainx-v5 and ./brainx-v5-cli also work as aliases for the main wrapper.

Advisory System (Pre-Action Check)

BrainX includes an advisory system that queries relevant memories, trajectories, and recurring patterns before executing high-risk tools. Helps agents avoid repeating past mistakes.

High-Risk Tools

The following tools automatically trigger advisory checks: exec, deploy, railway, delete, rm, drop, git push, git force-push, migration, cron, message send, email send.

CLI Usage

# Check for advisories before a tool execution
./brainx-v5 advisory --tool exec --args '{"command":"rm -rf /tmp/old"}' --agent coder --json

# Quick check via helper script
./scripts/advisory-check.sh exec '{"command":"rm -rf /tmp/old"}' coder

Agent Integration (Manual)

Since only agent:bootstrap is supported as a hook event, agents should manually call brainx advisory before high-risk tools:

# In agent SKILL.md or AGENTS.md, add:
# Before exec/deploy/delete/migration, run:
cd ~/.openclaw/skills/brainx-v5 && ./scripts/advisory-check.sh <tool> '<args_json>' <agent>

The advisory returns relevant memories, similar past problem→solution paths, and recurring patterns with a confidence score. It's informational — never blocking.

Agent-Aware Hook Injection

The agent:bootstrap hook uses agent profiles (hook/agent-profiles.json) to customize memory injection per agent:

• coder: Boosts gotcha/error/learning memories; filters by infrastructure/code/deploy/github contexts; excludes notes
• writer: Boosts decision/learning; filters by content/seo/marketing; excludes errors
• monitor: Boosts gotcha/error; filters by infrastructure/health/monitoring
• echo: No filtering (default behavior)

Agents not listed in the profiles file get the default unfiltered injection. Edit hook/agent-profiles.json to add new agent profiles.

Cross-Agent Memory Sharing

The hook reserves ~30% of injection slots for cross-agent memories, ensuring each agent sees relevant learnings from other agents. The cross-agent-learning.js script tags high-importance memories for cross-agent visibility without creating duplicates.

Security & Trust

This skill is flagged with "suspicious patterns" by ClawHub's automated scanner. Here's what each pattern does and why it's necessary:

No secrets are stored in code. All credentials come from environment variables. No data leaves the system except embedding requests to OpenAI.

Notes

• Memories are stored with vector embeddings (1536 dimensions)
• Search uses cosine similarity
• inject is the most useful tool for giving context to LLMs
• Tier hot = fast access, cold/archive = long-term storage
• Memories are persistent in PostgreSQL (independent of OpenClaw)
• Auto-injection hook fires on every agent:bootstrap

Feature Status (Tables)

✅ All Operational

8/8 tables operational. Population scripts implemented 2026-03-06.

Full Feature Inventory (35)

CLI Core (brainx <cmd>)

Processing Scripts (scripts/)

Hooks and Infrastructure

V5 Metadata

• sourceKind — Origin: user_explicit, agent_inference, tool_verified, llm_distilled, etc.
• sourcePath — Source file/URL
• confidence — Score 0-1
• expiresAt — Automatic expiration
• sensitivity — normal/sensitive/restricted
• Automatic PII scrubbing (BRAINX_PII_SCRUB_ENABLED)
• Similarity-based dedup (BRAINX_DEDUPE_SIM_THRESHOLD)