name: memory-curator description: Manage architectural decisions and insights in memory.jsonl. Use when you need to document strategic decisions, lessons learned, fixed problems, or architectural insights.

<memory_curator>

Memory Curator Skill

memory.jsonl stores durable engineering knowledge that can't be inferred from code.

• Never write to the file manually. Always use the add_memory_entry.py script.

The Critical Test

Before adding ANY memory entry, ask:

If the fix is in the code, don't add to memory. The code IS the memory.

SHOULD Add to Memory

1. Architectural patterns that affect how future features are built

   • Cross-cutting decisions (auth strategy, error handling approach, state management)
   • Patterns that must be followed consistently across modules

2. Data model gotchas not obvious from code

   • "Plaid sandbox returns per-share cost_basis, production returns total"
   • "Artemis stores cost_basis per-share, multiply by quantity for total"

3. External API limitations requiring workarounds

   • "Polygon.io doesn't support hourly bars on this plan tier"
   • "CloudFlare Workers has no filesystem - use build-time bundling"

4. Cross-agent coordination patterns

   • How agents pass context to each other
   • What data format specialized agents expect

SHOULD NOT Add to Memory

1. One-time bug fixes → The fix is in the code

   • ❌ "Fixed Hermes returning strings instead of numbers" → Code now does parseFloat()
   • ❌ "Fixed double-counting bug in portfolio calculation" → Code is fixed

2. SDK/library quirks → Once code handles it, done

   • ❌ "AI SDK useChat id prop doesn't transmit to server" → Code passes it in body
   • ❌ "Drizzle sql<Date> returns strings" → Code wraps in new Date()

3. Agent behavior rules → Belong in operating manual

   • ❌ "Never commit without asking" → Put in agent-operating-manual.md
   • ❌ "Never skip plan steps" → Put in agent-operating-manual.md

4. Module-specific implementation details → Belong in CLAUDE.md

   • ❌ "This service uses connection pooling" → Document in module's CLAUDE.md

Good Examples

• "Sequential agent pattern: tool-using agent first (mode:'generate'), then structured output agent receives results via promptVariables. Required because generateObject() doesn't support tools."
• "Portfolio validation must calculate ALL requirements before checking sufficiency. Two-pass approach: first calculate costs, then validate. Otherwise shows $0 transfer needed when insufficient."
• "LLM agents ignore validation tool errors unless prompt explicitly says what to do when valid=false. Must include iteration pattern with fix-and-retry loop."

Poor Examples (don't create)

• "Fixed the parseFloat bug in price service" → Code is fixed, no need to remember
• "Hermes API returns strings not numbers" → Code handles it, done
• "Used React Query for data fetching" → Obvious, no decision rationale
• "Implemented TASK-067" → Task summary belongs in task files

Workflow

Always use the helper script and never edit the file by hand.

python3 .claude/skills/memory-curator/scripts/add_memory_entry.py \
  --summary "<see Summary Format below>" \
  --tags architecture,api,lessons-learned \
  --links "TASK-090 services/backend-api/src/stripe/stripe-service.ts"

The script auto-detects project root by walking up to find .claude/ and .meridian/ directories. Note: if python3 is failing, try using python instead.

The script will:

• Compute the next sequential ID (mem-0001, mem-0002, …)
• Add a UTC timestamp (YYYY-MM-DDTHH:MM:SSZ)
• Append a single JSON object as one line to .meridian/memory.jsonl
• Echo the written entry for confirmation

Edit an existing entry

python3 .claude/skills/memory-curator/scripts/edit_memory_entry.py \
  --id mem-0042 \
  --summary "<new summary>" \
  --tags architecture,api \
  --links "TASK-090 docs/design.md"

• Provide at least one field to change (--summary, --tags, --links).
• Tags/links flags replace the lists entirely; include the full set you want to keep.

Delete an entry

python3 .claude/skills/memory-curator/scripts/delete_memory_entry.py \
  --id mem-0042

• Only delete when an entry is clearly obsolete or incorrect.
• The script rewrites the file without that entry; there is no undo.

Summary Format (consistent & skimmable)

Write --summary as concise paragraph with maximum 2-3 sentences. If it doesn't fit, it's probably a design doc, not a memory entry—link to that doc and keep the summary short.

Tags

Prefer a few broad tags over many hyper‑specific ones.

Examples: architecture, lessons-learned, problem-solved, pattern, decision, tradeoff, deprecation, migration

Links (make future retrieval easy)

Use --links for TASK IDs, critical file paths, or design docs. Examples:

--links "TASK-091 services/web/app/(auth)/route.ts docs/design-docs/authentication.md"

It should be a single string with quotes around it. </memory_curator>