Learnings Write

/learnings-write — Capture a Durable Learning

Append a curated entry to learnings.md (L4 project, default) or ~/.claude/ai-skills/learnings.md (L5 user-global, with --global). Spawns the memory-curator agent for dedupe, PII filter, schema validation per learnings-schema.md.

When to use

• After solving a tricky bug: "Async state in React 18 needs startTransition for non-urgent updates"
• After a design decision: "We chose gRPC over REST because client SDKs are lower-friction"
• After a pattern emerges: "Repository pattern works better than direct ORM for our team"
• After a gotcha: "Postgres text column ignores trailing whitespace in indexes"

Invocation

/learnings-write "Async state management — startTransition for non-urgent updates"
/learnings-write "Why we chose gRPC over REST" --global
/learnings-write "Postgres text column index trailing whitespace gotcha" --tags "postgres,index"
/learnings-write "ADR: adopt outbox pattern for cross-service events" --committed

Arguments

Behavior

1. Validate --global permission per memory-discipline.md rule 3 — must have BOTH userConfig.user_global_memory_enabled: true AND --global flag.
2. Spawn memory-curator agent (the agent is spawn-only — not user-invocable) with G7 spawn payload:
   • goal: "Add a learning entry the user wrote: "
   • state_slice: existing learnings.md (target layer) for dedupe context
   • allowed_tools: Read, Write
3. Memory-curator:
   • Reads existing learnings.md (target layer) for dedupe per memory-validation.md
   • Applies L5 scope check if --global (rejects entries with project paths/names per memory-discipline.md rule 4)
   • Applies PII filter (apply_pii_filter() from _lib.py)
   • Formats per learnings-schema.md (H2 entity heading + Type/Source/Confidence/Created/Last confirmed/Scope frontmatter + body)
   • Appends to target file (append-only)
4. Receive G7 return contract:
   • status: ok → entry written; report path + entity heading
   • status: partial → some content rejected (e.g., L5 scope violation); report what was rejected and why
   • status: needs_clarification → conflict with existing entry; surface the question to user

Hard rules

• Spawn-only memory-curator — never bypass and write learnings.md directly per memory-discipline.md rule 1
• PII filter mandatory — every write passes through filter; redactions logged to redactions.log
• .committed/ writes route through pre-tool-use-committed-write.py hook — committed learnings must match the allowlist pattern (default: learnings.md, architecture-decisions/*.md, etc.)
• L5 strict scope — entries with project-specific paths or names are rejected (must be generalizable patterns)

Failure modes

• --global without permission: refuse with same error as /memory-recall
• Entry too short (< 20 chars): refuse — entries should be substantive
• Entry duplicates existing: memory-curator returns status: needs_clarification with the existing entry; user picks merge / skip / supersede
• PII filter unavailable (_lib.py import error): memory-curator returns status: failed; never write without filter

Memory writes

Integration

• Spawns: memory-curator agent (Read, Write only; no Task; no Bash)
• Schemas: plugin/schemas/spawn-payload.schema.json, plugin/schemas/return-contract.schema.json
• Templates: plugin/memory/templates/learnings-schema.md (entry format), plugin/memory/templates/committed-allowlist.txt (committed write rules)
• Rules: memory-discipline (write rules + PII filter mandate + L5 scope), memory-validation (dedupe + conflict resolution)
• Hooks: pre-tool-use-committed-write.py (committed allowlist enforcement)
• Companion: /memory-recall to search; /memory-init to initialize the tree
• Layer model: plugin/docs/concepts/memory.md documents the L0–L5 layers (L4 = project memory, L5 = user-global). This skill writes only to L4 (default) and L5 (opt-in).