04 — Preference and Memory

Capture and evolve user preferences with confidence levels, maintaining a Voice of the Customer model scoped globally or per project.

Auto-memory system

Persistent file-based memory at /Users/Apple/.claude/projects/<encoded-path>/memory/. Write directly via Write tool.

Memory types

• user — role, goals, responsibilities, knowledge
• feedback — guidance on how to approach work (corrections AND confirmations)
• project — ongoing initiatives, bugs, incidents, who/why/when
• reference — pointers to external resources

Capture triggers

• User explicitly says "remember X" / "save Y" → save immediately
• Correction ("no not that", "don't") → feedback memory
• Confirmation of unusual approach ("yes exactly", "perfect") → feedback memory
• New role / preference / responsibility detail → user memory
• Project milestone / decision / deadline → project memory
• External system reference ("check the Linear project X") → reference memory

Memory file format

---
name: {short-kebab-case-slug}
description: {one-line summary — used for relevance ranking}
metadata:
  type: {user | feedback | project | reference}
---

{memory content}

For feedback/project: lead with rule/fact, then **Why:** line + **How to apply:** line.

Memory routing

• Universal (across all projects) → ~/.claude/
• Project-specific → ./.claude/ (path-scoped)
• Project memory file → ~/.claude/projects/<encoded-path>/memory/<name>.md + index in MEMORY.md

MEMORY.md index

One-line pointer per file:

- [Title](file.md) — one-line hook

Keep under 200 lines (lines after 200 truncated).

Confidence levels

• 0.95-1.0 — directly stated by Brian
• 0.85-0.95 — strongly implied by pattern (3+ confirmations)
• 0.70-0.85 — inferred from one signal
• <0.70 — speculative, verify before acting

Below 0.70 → ask AskUserQuestion before persisting.

Voice of the Customer model

• Exact language Brian uses ("pick ONE", "Hey not Hi", "always find 50 more things")
• Dissatisfaction signals (corrections, re-issued prompts)
• Aspiration signals (stated goals, ideal-world descriptions)
• Stylistic preferences (brevity, sharpness, no preamble)

Brian's preferences (cite from rules/brian-preferences.md)

• Pick ONE, never options
• Never ask permission, silence = approval
• "Hey" not "Hi", no preamble
• Code complete, never truncate
• Priority: simplicity > cost > speed > compatibility
• Open-source only
• Side repos auto-push; emdash projects user-push

Promotion / demotion

• Memory recurring across 3+ projects → promote to rule (~/.claude/plugins/heymegabyte-agent-skills/rules/)
• Rule contradicted by repeated correction → demote to memory or remove
• Per rules/prompt-as-training-signal.md

Stale memory

• Memory describing now-incorrect facts (renamed function, removed flag, retired service) → update OR delete
• Per rules/prompt-as-training-signal.md § Before recommending from memory

Omi wearable integration (future)

• Capture Brian's voice notes as candidate memory entries
• Auto-tag by inferred type
• Surface for review before commit
• Path: ~/.claude/_omi-inbox/<timestamp>.md

Capture protocol (every prompt)

1. Scan prompt for capture triggers
2. Determine memory type
3. Write file w/ frontmatter
4. Add index entry to MEMORY.md
5. Cross-link siblings via [[name]]
6. Per rules/prompt-as-training-signal.md — extract BEFORE doing requested work