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#

text

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#

bash

# 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.

markdown

# 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#

ToolDescriptionquery_graphSearch primary FalkorDB graph for related factssearch_webPerplexity AI web searchread_fileRead workspace .md fileslist_filesList workspace .md filessandbox_runRun Python snippets (numpy/scipy, no network, 30s limit)check_dead_endsCheck wander graph for previously closed threadsrecord_dead_endRecord a closed thread (lower bar than elevate)elevateWrite finding to MENTAL_EXPLORATION.md (strict gate)

Configuration#

Edit mind-wander/mind_wander_config.py:

python

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#

FileContentsIn graph-rag?MENTAL_EXPLORATION.mdElevated findings✅ via memwatchdDEAD_ENDS.mdClosed threads summary❌ nevercompletions/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.