Mind-Wander

Background reasoning agent that autonomously explores open questions using a local LLM (Qwen3.5-9B), a private knowledge graph for dead-end tracking, and Perplexity web search. Fires on a schedule, picks one unresolved question from ON_YOUR_MIND.md, runs sandbox experiments and web searches, and writes findings to MENTAL_EXPLORATION.md only when genuinely novel — mirroring hippocampal background consolidation. Uses a separate FalkorDB 'wander' graph so exploration history never pollutes the primary agent context. Use when: setting up autonomous background research for an OpenClaw agent, exploring research questions without consuming primary LLM token budget, building training data from exploratory reasoning sessions, or tracking dead ends to avoid re-exploration. Triggers on: "mind wander", "background reasoning", "autonomous research", "wander agent", "ON_YOUR_MIND", "MENTAL_EXPLORATION", "dead ends", "explore while I sleep".

Mind-Wander Skill

Autonomous background reasoning agent. Runs locally on Qwen3.5-9B, consumes zero Anthropic tokens, and elevates findings to your context only when genuinely novel.

How it works

ON_YOUR_MIND.md  →  Qwen3.5 wander agent (every 30min)
  open questions       ↓ tools: query_graph, search_web,
  tangents             ↓         read_file, sandbox_run,
  hypotheses           ↓         check_dead_ends, record_dead_end, elevate()
                       ↓
              novelty gate (strict)
                  ↙         ↘
           MENTAL_EXPLORATION.md    DEAD_ENDS.md + wander graph
           (elevated findings)      (closed threads, never in your context)
                  ↓
           memwatchd detects write
                  ↓
           graph-rag memory (your context)

Prerequisites

• Qwen3.5-9B-Q8 pulled to Ollama: ollama pull qwen3.5-wander-q8 (or use install script)
• FalkorDB running (shared with graph-rag-memory skill if installed)
• Perplexity API key (optional but recommended for web search)
• graph-rag-memory skill installed (recommended — shares FalkorDB and Ollama)

Quick start

# Install and set up
bash mind-wander/scripts/install.sh

# Write your first open question
echo "## What is the best approach to X?" >> ON_YOUR_MIND.md

# Run manually
python3 mind-wander/run.py --verbose

# Check findings
cat MENTAL_EXPLORATION.md
cat DEAD_ENDS.md

# Status
python3 mind-wander/run.py --status

The ON_YOUR_MIND.md anchor file

Create ON_YOUR_MIND.md in your workspace root with questions and tangents. The agent picks ONE per session. Format freely — the agent reads it as-is.

# On My Mind

## Open Questions
- Does X actually work better than Y in production?
- Is there a paper on Z that I haven't found yet?

## Tangents
- The implementation of A might connect to B in an interesting way

Mark completed items with ## ✅ COMPLETED so the agent skips them.

The novelty gate

• Restatement of known facts → discarded, nothing written
• New external finding intersecting open question → elevate()
• Empirical sandbox result that changes understanding → elevate()
• Definitively closed thread (≥2 targeted searches) → record_dead_end()

Tools available to the wander agent

Tool	Description
query_graph	Search primary FalkorDB graph for related facts
search_web	Perplexity AI web search
read_file	Read workspace .md files
list_files	List workspace .md files
sandbox_run	Run Python snippets (numpy/scipy, no network, 30s limit)
check_dead_ends	Check wander graph for previously closed threads
record_dead_end	Record a closed thread (lower bar than elevate)
elevate	Write finding to MENTAL_EXPLORATION.md (strict gate)

Configuration

Edit mind-wander/mind_wander_config.py:

WANDER_MODEL    = "qwen3.5-wander-q8"   # or q4 for lighter
WANDER_OLLAMA   = "http://172.18.0.1:11436"
MAX_TOOL_CALLS  = 20
COOLDOWN_HOURS  = 3   # min hours before revisiting same anchor item

Output files

File	Contents	In graph-rag?
MENTAL_EXPLORATION.md	Elevated findings	✅ via memwatchd
DEAD_ENDS.md	Closed threads summary	❌ never
completions/wander/	Full session JSON	❌ training data only

Research context

This skill produced the first novel finding in its 10-minute test run: "Cross-space routing (routing in nomic-space, retrieving in arctic/bge-m3 space) matches same-space baseline accuracy — suggesting domain routing is robust to embedding space discontinuities." See NOVELTY_LOG.md for tracked findings.

See references/research.md for theoretical foundations and references/setup.md for detailed installation instructions.