feat: mem:brainstorm -- rapid-fire knowledge graph population via web traversal

[robotrocketscience]

Concept

A new mem:brainstorm skill that rapidly populates the knowledge graph through automated web research and Wikipedia link traversal, then ingests everything into agentmemory for dynamic belief formation.

Envisioned Workflow

1. Prior art search: Web searches for the topic/hypothesis to find existing work
2. Wikipedia seeding: Find Wikipedia articles covering key topics
3. Link traversal game: Starting from a Wikipedia page, follow interesting links recursively:
   • Scan the article for key concepts
   • Click interesting links to related articles
   • Scan those articles, find more links
   • Backtrack or keep going based on relevance
   • Humans do this for hours and retain a fraction; LLM + agentmemory could do it in seconds and retain everything
4. Bulk ingestion: Dump all discovered information into the knowledge graph as uncertain beliefs
5. Dynamic refinement: Talk through findings, run tests, let agentmemory hooks start dynamically firing and connecting things -- the graph self-organizes as beliefs get used, confirmed, or contradicted

Key Insight

The value is in the combination: rapid uncertain ingestion + agentmemory's existing feedback loops. You don't need high-confidence input if you have a good refinement mechanism. The brainstorm populates the graph with raw material; normal usage shapes it into knowledge.

Technical Considerations

• Subagent parallelization for multiple search/traverse paths simultaneously
• TTL on brainstorm-sourced beliefs (like wonder_generated) so stale unvalidated ones decay
• Rate limiting on web requests (Wikipedia has usage policies)
• Source attribution (which URL, which article, which link path)
• Depth limits on traversal to prevent runaway exploration
• Integration with existing wonder/wonder_ingest infrastructure

Related

• mem:wonder (deep-dive research on specific hypotheses -- brainstorm is broader/shallower)
• wonder_ingest (speculative belief ingestion with TTL -- reusable for brainstorm output)
• feat: subagents have no access to agentmemory (MCP, hooks, injections) #35 (subagent memory -- brainstorm subagents would benefit from belief injection)

[robotrocketscience]

brainstorm, wonder, and reason should all be orthogonal mental models / metaphors of how to interact with the knowledge graph in a way that humans can relate to, since we are avoiding human-based memory models intentionally. but we still need interfaces via commands like wonder, reason, and now brainstorm as well as eventually the obsidian visualization layer, are ways to bridge the gap from human-incomprehensible super complicated graph with typed nodes and edges to natural invocations and usage and easily-mapped metaphors