Knowledge Loop in Action

Watch CQ evolve over 5 sessions — from knowing nothing about you to anticipating your preferences.

---

What is the Knowledge Loop?

Most AI tools start every session from zero. CQ closes the loop:

Session → Preferences captured → Rules generated → AI behavior changes

This isn't memory. It's evolution. Here's what it looks like in practice.

---

Session 1: You Explain Everything

You're working on a mesh recovery research project.

You: "Run the experiment with MPJPE as primary metric.
      Always check MPJPE before looking at other metrics."

CQ: runs experiment, reports MPJPE first

When the session ends, CQ captures:

Preference detected (count: 1):
  "Check MPJPE metric first when evaluating experiments"
  Level: observation

Nothing changes yet. One mention isn't a pattern.

---

Session 2: You Mention It Again

You: "What's the MPJPE? Show me that first."

Preference reinforced (count: 2):
  "Check MPJPE metric first"
  Level: observation

Still just tracking. Two mentions could be coincidence.

---

Session 3: Pattern Emerges

You: "Start with MPJPE numbers, then we'll look at the rest."

Preference confirmed (count: 3):
  "Check MPJPE metric first"
  Level: hint → written to CLAUDE.md

CQ writes a hint to your project's CLAUDE.md:

# Hints (auto-generated from session patterns)
- Check MPJPE metric first when evaluating experiment results

Next session, this hint is loaded into the AI's system prompt.

---

Session 4-5: Becomes a Rule

By session 5, the same preference has appeared 5 times:

Preference promoted (count: 5):
  "Check MPJPE metric first"
  Level: rule → written to .claude/rules/

CQ creates a rule file:

# .claude/rules/experiment-metrics.md
- Always report MPJPE as the primary metric in experiment results
- Show MPJPE before PA-MPJPE, HD, or other secondary metrics

Rules are stronger than hints — they're loaded into every session's system prompt.

---

Session 6+: CQ Already Knows

CQ: "Experiment complete. Results:
     MPJPE: 45.2mm (↓3.1 from baseline)
     PA-MPJPE: 38.7mm
     HD: 1.74mm"

You didn't ask. CQ already knows MPJPE comes first.

---

Real Example: 5 Research Sessions

After 5 mesh recovery research sessions, CQ auto-generated these patterns:

Count	Level	What CQ Learned
5x	Rule	"Run experiments via Hub automatically"
4x	Hint	"Use @key=value metric output format"
4x	Hint	"Check MPJPE/HD/MSD metrics first"
3x	Hint	"Single-DRR experiments before multi-view"

---

How Preferences Flow

You work normally
       │
       ▼
Session ends → CQ captures decisions, preferences, discoveries
       │
       ▼
Count < 3: stored as observation (invisible)
Count = 3: promoted to hint (CLAUDE.md)
Count = 5: promoted to rule (.claude/rules/)
       │
       ▼
Next session: AI loads rules + hints into system prompt
       │
       ▼
AI behavior changes — without you asking

---

Managing Your Growth

See what CQ has learned

cat CLAUDE.md              # Hints
ls .claude/rules/          # Rules

Delete a rule you disagree with

rm .claude/rules/experiment-metrics.md

Deleted rules are permanently suppressed — CQ won't re-generate them.

Knowledge flows across AI tools

Preferences stored via Remote MCP are available everywhere:

• Learned in Claude Code → available in ChatGPT
• Learned in Cursor → available in Claude Desktop

One brain. Consistent behavior. Everywhere.

---

Next Steps

• ChatGPT → Claude — cross-AI knowledge flow in practice
• Research Loop — automated experiment cycles