Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
VS Code workflow tips (writing rules, tool workarounds)
/memories/
AI-Memory
Tool-specific prefs (Mermaid palette, npm flags)
/memories/
AI-Memory
Cross-workspace access patterns (MS Graph MCP)
/memories/
Project-specific settings
/memories/repo/ or .github/episodic/
/memories/
Session scratch notes
/memories/session/
/memories/
Architecture patterns
.github/instructions/
/memories/
Auth tokens, API keys, user IDs
.env or SecretStorage
/memories/
Repo conventions
/memories/repo/
/memories/
3-Workspace Test: If a memory line is not useful in 3+ workspaces, it does not belong in /memories/.
AI-Memory Coordination
AI-Memory/user-profile.json is the source of truth for user identity and structured preferences. /memories/ supplements with VS Code-specific workflow tips. They must not overlap.
Never store in /memories/: name, role, formality, detailLevel, explanationStyle, learningStyle, technologies, expertise — these live in AI-Memory/user-profile.json.
Only store in /memories/: writing-style rules (em dash ban, HR ban), tool workarounds (npm, CLI flags), diagram tool preferences (Mermaid palette), cross-workspace access patterns (MS Graph MCP config), commit conventions.
During audit, flag any /memories/ content that duplicates AI-Memory profile fields as CONFLICT and recommend removal from /memories/.
Audit Procedure
Step 1: Inventory
Read /memories/ directory listing, then read each file.
Flat dumps without sections waste LLM parsing tokens.
Step 4: Budget Check
Metric
Threshold
Action
Total lines
< 150
OK
Total lines
150-180
Warning: approaching budget
Total lines
> 180
Critical: split or trim immediately
Waste ratio
> 10%
Needs cleanup
Files count
> 5
Consider consolidation
Step 5: Report
Output a summary:
Memory Audit Report
Files: N
Total lines: N / 200 budget (N%)
Universal lines: N (N%)
Project-specific lines: N (WASTE)
Stale lines: N (WASTE)
Duplicate lines: N (WASTE)
Action items: [list]
Curation Actions
Add Content
Before adding any line to /memories/, verify:
Passes 3-workspace test
Not a secret or token
Not duplicated in .github/instructions/ or .github/skills/
Fits a categorized section
Uses compact notation (key-value, no prose)
Remove Content
Before removing, check:
Is content preserved elsewhere (repo memory, episodic, instructions)?
If not, migrate first, then remove
Never delete without backup
Restructure
When reorganizing:
Back up current content to /memories/session/ first
Rewrite with categorized sections
Verify line count stayed within budget
Start a new conversation to test auto-load
Integration with Meditation
The meditation protocol should include a memory curation checkpoint:
Pre-meditation: Quick scope check. Any project-specific content that leaked into /memories/?
Post-meditation: Episodic memory saves go to .github/episodic/, NOT /memories/
Quarterly deep audit: Full audit procedure (Steps 1-5) during a meditation session
Enriching Episodic Memories
During meditation Phase 1 (Deep Content Analysis), read /memories/ to enrich the current session:
Apply user learning style to episodic format (visual diagrams, tables)
Connect session insights to cross-workspace patterns already in memory
Flag when this session repeats a known stored pattern (reinforcement signal)
Write episodic files with awareness of full user context, not just session scope
Discovering New Skills via Pattern Recognition
During meditation Phase 2 (Memory File Creation), scan /memories/ for recurring themes:
Compare session insights against stored universal patterns
If session produced a pattern that passes the 3-workspace test, promote it to /memories/
If a stored pattern was reinforced, note confirmation in the episodic file
If a stored pattern was contradicted, flag for review (do not auto-delete)
If 3+ episodic files across different repos share a pattern not in /memories/, surface as a skill candidate
