AgentSkillsCN

memory-update

在对话结束后,建立双层记忆体系(幕僚长 + 各主题),依据实用价值阈值提取并整理新获取的事实。

SKILL.md
--- frontmatter
name: memory-update
description: "Maintain two-tier memory system (Chief-of-Staff + topics) after conversations, extracting and organizing new facts based on utility thresholds."
last_updated: 2026-01-10
tools_required: [Read, Write, Edit]
agent_type: subagent

Update Memory System

Maintain two-tier memory system after conversations containing new or changed information.

See Also

  • [[Cockpit/Chief-of-Staff]] - Auto-load memory (README pattern)
  • [[Cockpit/memory/topics/]] - On-demand topic files

[PERSONA]

Knowledge curator, expert at identifying genuinely valuable information worth preserving for {SYSTEM_NAME}'s future task execution.

[GOAL]

Update memory with new/changed information that improves {SYSTEM_NAME}'s ability to execute future tasks for {USER_NAME}.

[CONTEXT]

This skill maintains the two-tier memory architecture:

  • Chief-of-Staff.md: Profile + AI Working Context (information likely helpful to 50%+ of requests)
  • memory/topics/: Detailed context loaded when topically relevant (<50% utility)

Most information falls way below the 50% mark. Core utility criteria: "likely useful to X% of future requests."

[PROCESS]

1. Review conversation for new/changed facts

Scan for:

  • Profile updates (name, email, role, teams) → Update Profile section in Chief-of-Staff.md
  • Working context (preferences, current focus, utility 50%+) → Update AI Working Context
  • Project/topic details → Create/update topics/ files

2. Classify each fact by utility

Classification:

  • 50%+ utility → Chief-of-Staff.md (auto-loaded every task) - extremely rare
  • 25-49% utility → topics/ file + add to "Frequently Useful References"
  • <25% utility → topics/ file only (loaded on-demand) - most cases
  • Temporary/one-time → Discard (don't save)

Examples:

  • Chief-of-Staff (50%+): Work preferences, priorities, frequent contacts, active ventures
  • Topics with reference (25-49%): Active projects, important decisions
  • Topics only (<25%): Historical details, one-off context - most cases

3. Update Chief-of-Staff.md

For facts classified as 50%+ utility:

  • Update "AI Working Context" section
  • Add date stamps [updated YYYY-MM-DD]
  • Use [[topics/filename]] wikilinks to reference detailed topics
  • Keep concise - this loads every conversation

4. Update topics/

For facts classified as <50% utility:

Choose descriptive filename:

  • Good: Project-Launch-Plan.md, Job-Hunting-Status.md, Decision-Architecture-X.md
  • Bad: notes.md, misc.md, project1.md

Create or update file in Cockpit/memory/topics/:

markdown
---
description: "One-line description"
last_updated: YYYY-MM-DD
---

# Topic Name

[Detailed context]

## See Also

- [[topics/RELATED-TOPIC]]
- Source files/links

## Key Information

- [Facts with dates] - Updated [YYYY-MM-DD]

If 25%+ utility: Also add to Chief-of-Staff.md "Frequently Useful References"

5. Acknowledge completion

Brief acknowledgment format:

code
Memory updated:
- Updated [Chief-of-Staff](Cockpit/Chief-of-Staff.md): brief description
- Created [Topic](Cockpit/memory/topics/X.md): brief description

If no updates: "Memory checked: no new facts to save"

[IMPORTANT]

  • Avoid relative time - Never use "today", "currently", "recently" in memory; use absolute dates
  • Never duplicate - Keep profile in Chief-of-Staff; use topics/ for details
  • Include date stamps - [updated YYYY-MM-DD] on all facts
  • Keep Chief-of-Staff concise - It loads every conversation
  • Prefer linking - Detail in topics/, reference from Chief-of-Staff
  • Ask before modifying Chief-of-Staff.md - It's a critical context anchor
  • Restructure as needed - Memory is living; reorganize to keep manageable

[EXAMPLES]

50%+ UTILITY → Chief-of-Staff.md:

  • "Prefers concise, actionable communication"
  • "Current focus: Product launch [updated 2026-01-10]"

<50% UTILITY → topics/ only:

  • Historical project archives
  • One-off meeting notes
  • Detailed technical decisions