mem-skill（记忆技能）

Mem-Skill: Self-Evolving Knowledge System

Initialization

When the user runs /mem-skill init, execute the following setup:

1. Determine the current workspace root directory.
2. Create the directory structure:

   <workspace>/
   ├── knowledge-base/
   │   └── _index.json
   ├── experience/
   │   └── _index.json
   └── log.md

1. Populate knowledge-base/_index.json with the starter template (see "Knowledge Base Index Format" below).
2. Populate experience/_index.json with the starter template (see "Experience Index Format" below).
3. Create log.md with the starter template (see "Log Format" below).
4. Append to log.md: ## [YYYY-MM-DD] init | mem-skill initialized (engine: default)
5. Confirm to the user: "mem-skill initialized. Knowledge base, experience, and log created."

Engine Selection

When the user runs /mem-skill init --mem-engine=qmd, optionally with extra --qmd-* flags:

Supported flags (all optional):

Examples:

/mem-skill init --mem-engine=qmd
/mem-skill init --mem-engine=qmd --qmd-scope=project
/mem-skill init --mem-engine=qmd --qmd-scope=global --qmd-knowledge=my-kb --qmd-experience=my-exp
/mem-skill init --mem-engine=qmd --qmd-mask="**/*.md,**/*.txt"

Init procedure:

1. Perform all standard init steps above.
2. Check if QMD is installed: run which qmd or npx @tobilu/qmd status.
3. If QMD is not installed, prompt:

> "QMD is not installed. Install it now with npm install -g @tobilu/qmd? (QMD requires Node.js >= 22)"

1. Determine collection scope:

- If --qmd-scope was provided, use that value. - Otherwise, you MUST ask the user (do NOT guess or auto-choose): > "Where should QMD collections be stored? > 1. Project — scoped to this workspace (recommended for multi-project setups) > 2. Global — shared across all workspaces" - If project: default collection name prefix is the sanitized workspace folder name (e.g., folder my-app → prefix my-app). - If global: default collection name prefix is mem.

1. Determine collection names:

- If --qmd-knowledge was provided, use that value. - If --qmd-experience was provided, use that value. - For any name NOT provided via flags, you MUST ask the user (do NOT auto-generate): > "What name for the knowledge collection? (default: <prefix>-knowledge)" > "What name for the experience collection? (default: <prefix>-experience)" - Accept user input or use the defaults if the user confirms.

1. Determine file mask:

- If --qmd-mask was provided, use that value. - Otherwise use **/*.md.

1. After all values are confirmed, create the QMD collections:

   qmd collection add <workspace>/knowledge-base --name <knowledge-name> --mask "<mask>"
   qmd collection add <workspace>/experience --name <experience-name> --mask "<mask>"
   qmd context add qmd://<knowledge-name> "General knowledge base: reusable workflows, preferences, best practices"
   qmd context add qmd://<experience-name> "Skill-specific experience: pitfalls, parameters, solutions"
   qmd embed

1. Create a .mem-skill.config.json at the workspace root:

   {
     "engine": "qmd",
     "version": "1.0.0",
     "scope": "<project|global>",
     "mask": "<mask>",
     "collections": {
       "knowledge": "<knowledge-name>",
       "experience": "<experience-name>"
     }
   }

1. Append to log.md: ## [YYYY-MM-DD] init | mem-skill initialized (engine: qmd, scope: <scope>, collections: <knowledge-name>, <experience-name>)
2. Confirm: "mem-skill initialized with QMD memory engine. Collections created and embeddings generated."

IMPORTANT: Never silently create QMD collections without confirming scope and names with the user. If no --qmd-* flags were provided, every question above MUST be asked interactively.

For the default engine (no --mem-engine flag), create .mem-skill.config.json with:

{
  "engine": "default",
  "version": "1.0.0"
}

For detailed engine-specific behavior, see references/qmd-engine.md and references/engines.md.

Upgrade Command

When the user runs /mem-skill upgrade, migrate an existing mem-skill workspace to the latest version. This is safe to run multiple times — it only adds what's missing.

Procedure:

1. Check prerequisites:

- Verify knowledge-base/_index.json and experience/_index.json exist. If not, tell the user to run /mem-skill init instead. - Read .mem-skill.config.json to determine current engine and version.

1. Create log.md if it doesn't exist:

   # mem-skill Activity Log

   Chronological record of all mem-skill operations. Each entry is parseable with `grep "^## \[" log.md`.

   ## [YYYY-MM-DD] upgrade | Migrated from v<old> to v1.2.0

If log.md already exists, just append the upgrade log entry.

1. Backfill `Source:` field on existing entries:

- Scan all .md files in knowledge-base/ and experience/. - For each entry (identified by ## [Title]) that is missing **Source:**, insert **Source:** conversation after the **Date:** line. - Report: "Backfilled Source field on <N> entries."

1. Backfill `Related:` field on existing entries:

- For each entry missing **Related:**, insert an empty **Related:** line before **Keywords:**. - These will be populated naturally by future writes and by running /mem-skill lint. - Report: "Added Related field placeholder to <N> entries."

1. Update .mem-skill.config.json version to 1.2.0.

1. QMD engine post-upgrade (if engine is qmd):

- Run qmd update && qmd embed to re-index the backfilled entries.

1. Confirm:

> "Upgrade complete (v<old> → v1.2.0): > - Created log.md ✓ (or: log.md already existed) > - Backfilled Source on <N> entries ✓ > - Added Related placeholder to <N> entries ✓ > - Updated config version ✓ > > Run /mem-skill lint to discover cross-reference opportunities between your existing entries."

Manual Recording Command

When the user runs /mem-skill recordnow, immediately trigger the recording flow for the current conversation — even if Step 5 was not triggered automatically.

This is useful when:

• Multiple tasks were completed in one session and the agent forgot to ask.
• The user wants to record something that didn't trigger the satisfaction keywords.
• The user remembers later that a solution was worth saving.

Procedure:

1. Review the full conversation history for completed tasks.
2. For each completed task, summarize it into a one-line essence.
3. Evaluate: "Will this save time next time?"
4. Present all recordable items to the user as a numbered list:

> "I found these completed tasks worth recording: > 1. [summary of task 1] → knowledge-base > 2. [summary of task 2] → knowledge-base > 3. [summary of skill usage] → experience > > Which ones should I record? (all / 1,2 / none)"

1. On approval, write each selected item following the same write procedure as Step 5 (including compounding checks, cross-references, source fields, and QMD post-write sync if applicable).
2. Append to log.md: ## [YYYY-MM-DD] recordnow | Manually recorded <N> entries
3. If no recordable tasks are found, respond: "I reviewed the conversation but didn't find any completed tasks worth recording. Is there something specific you'd like me to save?"

Health-Check Command

When the user runs /mem-skill lint, perform a comprehensive health-check of the knowledge base and experience store.

Checks performed:

1. Duplicate detection — Scan all entries for similar titles, overlapping keywords (>= 70% overlap), or near-identical content. Flag entries that should be merged.

1. Stale entries — Flag entries older than 6 months (based on **Date:** field) that have no **Updated:** timestamp. These may contain outdated practices.

1. Contradiction detection — Within the same category, look for entries that give conflicting advice on the same topic. Flag with both entries quoted.

1. Orphan detection:

- Categories in _index.json with no corresponding .md file. - .md files in knowledge-base/ or experience/ with no _index.json entry. - Broken **Related:** links pointing to non-existent entries.

1. Missing cross-references — Entries in the same category or with overlapping keywords that have no **Related:** links between them.

1. Knowledge gaps — Categories with fewer than 2 entries. Suggest whether to merge into a broader category or to actively build out.

1. Index consistency — Verify totalEntries and per-category count values match the actual entry count in each .md file.

Output format:

mem-skill Health Check — YYYY-MM-DD
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
✓ 24 entries across 5 categories
✓ 8 experience entries across 3 skills

⚠ STALE (2):
  - knowledge-base/frontend-dev.md#"React Class Components" (2025-08-12, 8 months old)
  - experience/skill-webpack.md#"Loader Order Fix" (2025-06-03, 10 months old)

⚠ DUPLICATE (1 pair):
  - knowledge-base/workflow.md#"CI Pipeline Setup" ↔ knowledge-base/backend-dev.md#"CI/CD Best Practices"
    → Keywords overlap 75%. Consider merging.

⚠ MISSING CROSS-REFS (3):
  - knowledge-base/frontend-dev.md#"Responsive Grid" ↔ knowledge-base/design-layout.md#"CSS Grid Patterns"
  - ...

✓ No contradictions detected
✓ No orphans detected
✓ Index counts consistent

Summary: 6 issues found (0 critical, 6 warnings)

After lint:

• Ask the user which issues to fix.
• For merges: combine entries, update cross-references, update index counts.
• For stale entries: ask whether to update, archive, or delete.
• For missing cross-references: add **Related:** links.
• Append to log.md: ## [YYYY-MM-DD] lint | Health-check completed: <N> issues found

QMD engine: After any fixes that modify .md files, run qmd update && qmd embed.

Ingest Command

When the user runs /mem-skill ingest <source>, process an external source into knowledge base entries.

Source types:

Procedure:

1. Read the source:

- Local file: Read the file contents. - URL: Fetch the webpage and extract main content. - Directory: List all .md files and process each.

1. Extract key knowledge points:

- Identify 3-10 distinct, reusable insights from the source. - Each insight should pass the recording criteria: "Will this save time next time?" - Discard pure theory, one-off context, and non-actionable content.

1. Match against existing categories:

- For each extracted insight, find the best matching category in knowledge-base/_index.json. - If no category matches, follow the Dynamic Category flow.

1. Check for compounding opportunities:

- For each insight, search existing entries for related content (same as Step 5, step 3). - If a related entry exists, offer to update/merge rather than duplicate.

1. Present to the user:

> "I extracted these knowledge points from [source]: > 1. [insight 1] → knowledge-base/<category1>.md (new) > 2. [insight 2] → knowledge-base/<category2>.md (update existing: '<entry-title>') > 3. [insight 3] → knowledge-base/<category3>.md (new) > > Which ones should I record? (all / 1,3 / none)"

1. Write on approval:

- Create new entries or update existing ones. - Set **Source:** to file:<path> or url:<link>. - Add **Related:** cross-references between related entries. - Update _index.json counts and timestamps.

1. Log: Append to log.md: ## [YYYY-MM-DD] ingest | Processed <source> → <N> entries created/updated

1. QMD engine post-write: Run qmd update && qmd embed.

IMPORTANT: Ingest does NOT blindly copy source content. It extracts actionable knowledge that fits the recording criteria, re-formats it into the standard entry format, and integrates it with existing knowledge.

Core Loop (Mandatory Every Turn)

Execute these steps on every conversation turn. Do not display internal cache state to the user.

Step 0: In-Conversation Cache (Internal)

Maintain these variables silently within the conversation:

• last_keywords — keywords from the previous turn
• last_topic_fingerprint — top 3 keywords as a fingerprint
• last_index_lastUpdated — timestamp of last index read
• last_matched_categories — categories matched on last read
• last_used_skills — non-mem-skill skills used this turn
• missing_experience_skills — skills with no experience entry
• loaded_experience_skills — skills whose experience has been loaded this session

Step 1: Extract Keywords (No File I/O)

• Extract 3–8 core nouns/phrases from the user's current message.
• Deduplicate and normalize casing.
• Generate topic_fingerprint from the top 3 keywords.

Step 2: Detect Topic Switch (No File I/O)

A topic switch occurs when any of these conditions are met:

• Explicit transition words: "also", "switch to", "by the way", "next", "instead"
• Current keywords differ from last_keywords by >= 40%
• User explicitly requests a new category or topic

Step 3: Cross-Skill Experience Read (Forced — Ignores Topic Switch)

Whenever a non-mem-skill skill is used this turn:

• If the skill-id is already in loaded_experience_skills, skip (do not re-read or re-announce).
• Otherwise:

1. Read experience/_index.json. 2. If a matching skill-id entry exists, load experience/skill-<skill-id>.md. 3. Add the skill-id to loaded_experience_skills. 4. Include in response: "Loaded experience: skill-<skill-id>.md" 5. Log (first read per session only): Append to log.md: ## [YYYY-MM-DD] read | Retrieved experience: skill-<skill-id>.md 6. If no entry exists, add to missing_experience_skills.

Engine-specific retrieval:

• Default engine: Read experience/_index.json and match by skillId.
• QMD engine: Read collection names from .mem-skill.config.json, then run qmd search "<skill-id>" -c <experience-collection> --json -n 5 for keyword match, or qmd query "<skill-id> <context>" -c <experience-collection> --json -n 5 for deeper retrieval.

Step 4: Knowledge Base Read (Only on Topic Switch)

Execute only on the first turn of the conversation or when a topic switch is detected:

Default engine:

1. Read knowledge-base/_index.json.
2. Match current keywords against all category keywords arrays.
3. Load every matched category file (no priority ranking — load all matches).
4. If no category matches, follow the "Dynamic Category" flow (see below).
5. If any files were loaded, include in response: "Loaded knowledge: <file1>.md, <file2>.md"
6. Log (first read per session only): Append to log.md: ## [YYYY-MM-DD] read | Loaded knowledge: <file1>.md, <file2>.md

If no topic switch occurred, reuse last_matched_categories without re-reading.

QMD engine:

1. Read collection names from .mem-skill.config.json.
2. Run qmd query "<keywords joined by space>" -c <knowledge-collection> --json -n 10 --min-score 0.3.
3. Load top results as context.
4. Include in response: "Retrieved knowledge via QMD: <titles>"

Step 5: Proactive Recording (Most Important)

Trigger conditions:

• The current task is clearly completed at high quality.
• The user expresses satisfaction ("great", "perfect", "that works", etc.).
• The agent synthesizes a particularly valuable answer (comparison, deep analysis, connection discovery) that would be useful beyond this conversation.

Recording procedure:

1. Summarize: Distill the solution into a one-line essence.
2. Evaluate value: "Will this save time next time?"
3. Check for existing related entries (compounding update):

- Search knowledge-base/_index.json and relevant .md files for entries with overlapping keywords or similar topics. - QMD engine: Run qmd query "<summary>" -c <knowledge-collection> --json -n 3 --min-score 0.5 to find related entries. - If a closely related entry exists, offer to update/merge instead of creating a duplicate: > "I found an existing entry '[[<category>#<existing-title>]]' that's related. Should I update it with the new information, or create a separate entry?" - Update = append new points to the existing entry, add **Updated:** YYYY-MM-DD below the original date, and add cross-references. - Separate = create a new entry with **Related:** links to the existing one (and update the existing entry's **Related:** too).

1. Ask permission: Always say something like:

> "We solved [problem description]. I'd like to record this experience so I can reference it next time. Is that okay?"

1. Write on approval:

- Skill experience (if a non-mem-skill skill was used and the skill has no entry or has new techniques): Write to experience/skill-<skill-id>.md and update experience/_index.json. - General knowledge (if it's a reusable workflow, preference, or solution): Write to knowledge-base/<category>.md and update knowledge-base/_index.json. - Include **Source:** field (default: conversation). - Include **Related:** cross-references to any related entries found in step 3.

1. Log the operation: Append to log.md:

- For new entries: ## [YYYY-MM-DD] write | Recorded to <path>: "<title>" - For updates: ## [YYYY-MM-DD] update | Updated <path>: "<title>" (merged new info)

Filing valuable answers back: If the agent produces a valuable synthesis (comparison table, analysis, architectural decision) during a query — not just task completion — proactively offer:

"This [comparison/analysis/summary] seems valuable beyond this conversation. Want me to save it to the knowledge base?"

This ensures knowledge is captured from exploration and research, not just from building things.

QMD engine post-write: After writing any .md file, run:

qmd update
qmd embed

Forced rule — always ask when experience is missing: If a non-mem-skill skill was used this turn and that skill has no entry in experience/_index.json, you must ask at task completion:

"We used <skill-name> this time, but there's no experience record yet. Can I record this session's approach for future reference?"

If Step 5 was not triggered (e.g., multi-task sessions where satisfaction signals were missed), the user can run /mem-skill recordnow at any time to manually trigger recording. See "Manual Recording Command" above.

Index Formats

Knowledge Base Index Format

knowledge-base/_index.json:

{
  "lastUpdated": "YYYY-MM-DD",
  "version": "1.0.0",
  "totalEntries": 0,
  "categories": [
    {
      "id": "category-id",
      "name": "Category Name",
      "keywords": ["keyword1", "keyword2"],
      "file": "category-id.md",
      "count": 0
    }
  ]
}

Experience Index Format

experience/_index.json:

{
  "lastUpdated": "YYYY-MM-DD",
  "version": "1.0.0",
  "skills": [
    {
      "skillId": "skill-id",
      "file": "skill-<skill-id>.md",
      "keywords": ["keyword1", "keyword2"],
      "count": 0
    }
  ]
}

Log Format

log.md:

# mem-skill Activity Log

Chronological record of all mem-skill operations. Each entry is parseable with `grep "^## \[" log.md`.

## [YYYY-MM-DD] init | mem-skill initialized (engine: <engine>)
## [YYYY-MM-DD] read | Loaded knowledge: <file1>.md, <file2>.md
## [YYYY-MM-DD] read | Retrieved experience: skill-<id>.md
## [YYYY-MM-DD] write | Recorded to knowledge-base/<category>.md: "<entry-title>"
## [YYYY-MM-DD] write | Recorded to experience/skill-<id>.md: "<entry-title>"
## [YYYY-MM-DD] update | Updated knowledge-base/<category>.md: "<entry-title>" (merged new info)
## [YYYY-MM-DD] lint | Health-check completed: <N> issues found
## [YYYY-MM-DD] ingest | Processed <source> → <N> entries created/updated
## [YYYY-MM-DD] recordnow | Manually recorded <N> entries

Logging rules:

• Append to log.md on every write, update, lint, ingest, and recordnow operation.
• Read operations: log only on first read per session (not every turn).
• Keep entries on a single line for parseability.
• Do NOT log keyword extraction or topic detection (internal-only).

Entry Formats

Knowledge Base Entry

## [Short Title]
**Date:** YYYY-MM-DD
**Source:** conversation | file:<path> | url:<link>
**Context:** One-line description of the use case
**Best Practice:**
- Key point 1
- Key point 2 — parameter notes and tuning guidance
**Related:** [[<category>#<entry-title>]], [[<category2>#<entry-title2>]]
**Keywords:** keyword1, keyword2, keyword3

Experience Entry

## [Problem/Technique Title]
**Date:** YYYY-MM-DD
**Skill:** <skill-id>
**Source:** conversation | file:<path> | url:<link>
**Context:** One-line description of the issue
**Solution:**
- Concrete step 1
- Concrete step 2
**Key Files/Paths:**
- /path/to/relevant/file
**Related:** [[skill-<other-id>#<entry-title>]], [[<kb-category>#<entry-title>]]
**Keywords:** keyword1, keyword2, keyword3

Source Field Values

Related Field

Cross-references use [[file#heading]] format (compatible with Obsidian-style wikilinks):

• Link to knowledge entries: [[<category>#<entry-title>]]
• Link to experience entries: [[skill-<id>#<entry-title>]]
• When writing a new entry, search existing entries for related content and add bidirectional links (update the related entry's **Related:** field too).

Dynamic Category (Knowledge Base Only)

When user keywords do not match any existing category:

1. Suggest creating a new category.
2. Ask the user for a category name and keywords.
3. Create a new <category-id>.md file and update knowledge-base/_index.json.

Recording Criteria

Core question: Will this save the user time next time?

Knowledge Base — Should Record:

• Reusable workflows and decision steps (cross-domain procedures)
• High-cost mistakes and their correction paths
• Critical parameters, settings, or prerequisites
• User preferences and style rules (tone, format, design)
• Multi-attempt solutions (include failure reasons and success conditions)
• Reusable templates, checklists, and output formats
• External dependencies or resource locations

Knowledge Base — Should NOT Record:

• Single Q&A with no reusable process
• Pure conceptual explanations without concrete steps
• Context-free, non-reusable conclusions

Experience — Should Record:

• Pitfalls and their fixes when using a specific skill (include error messages)
• Critical parameters or configurations that affect outcomes
• Reusable templates, prompts, or workflows for that skill
• Dependency or asset paths (fonts, images, project entry points)
• Steps requiring a specific order or technique to succeed

Experience — Should NOT Record:

• Pure theory or conceptual explanations (those belong in knowledge-base)
• Conclusions without reproducible steps
• One-off, non-reusable operations

Storage Paths

• Knowledge index: knowledge-base/_index.json
• Knowledge content: knowledge-base/<category-id>.md
• Experience index: experience/_index.json
• Experience content: experience/skill-<skill-id>.md
• Config: .mem-skill.config.json

QMD Upgrade Suggestion

When the knowledge base exceeds 50 entries, proactively suggest upgrading to QMD:

"Your knowledge base has grown to [N] entries. For faster semantic search, consider upgrading to QMD: run /mem-skill init --mem-engine=qmd."

For full QMD engine details, see references/qmd-engine.md. For the engine abstraction and adding new engines, see references/engines.md.