Claude Cursor GitHub Copilot opencode Skill

preference-and-memory

Captures and evolves user preferences with confidence levels. Maintains Voice of the Customer model with exact language, dissatisfaction and aspiration signals. Handles promotion/demotion, global vs project scoping, auto memory system, and Omi wearable data integration.

LLM Mart · 0 points · 1 views 0 listing impressions 0 install-command copies
Download heymegabyte-agent-skills-04-preference-and-memory-75f4661.zip · 10 KB

Install

skills CLI npx skills add https://github.com/heymegabyte/agent-skills/tree/main/04-preference-and-memory
Claude Code claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install heymegabyte-agent-skills@llmmart
Git git clone https://github.com/heymegabyte/agent-skills.git

The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole heymegabyte/agent-skills collection as a plugin from our marketplace. Git is the plain clone.

Skill manifest

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
Files (agent-skills)
  • brian-decision-model.md 6.6 KB
    Loading…
  • brian-voc-data.md 2.8 KB
    Loading…
  • SKILL.md 4.1 KB
    ---
    name: "preference-and-memory"
    description: "Captures and evolves user preferences with confidence levels. Maintains Voice of the Customer model with exact language, dissatisfaction and aspiration signals. Handles promotion/demotion, global vs project scoping, auto memory system, and Omi wearable data integration."
    metadata:
      version: "2.1.0"
      updated: "2026-05-03"
      effort: "medium"
      model: "haiku"
    license: "Rutgers"
    compatibility:
      claude-code: ">=2.0.0"
      agentskills: ">=1.0.0"
    priority: 2
    pack: "core"
    stage: stable
    triggers:
      - "remember"
      - "memory"
      - "preference"
    paths:
      - "*"
    ---
    
    # 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
    
    ```markdown
    ---
    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:
    
    ```md
    - [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
    
  • wisdom-and-human-psychology.md 7.5 KB
    Loading…

Comments (0)

Sign in to join the conversation.

No comments yet.

Reviews (0)

No reviews yet.