Extract learnings from audit history and conversation, manage memories, clean up stale entries. Use after completing a track, during retrospective (③ memory step), at end of session, or anytime the user wants memory maintenance. Triggers on 'what did we learn', 'memory cleanup', 'review learnings', 'retrospective', 'update memories', '회고', '메모리 정리'.

Retrospect Protocol

Mine the current session for learnings and translate them into durable memories. Sessions generate massive amounts of information — your job is to find the fraction worth remembering, structure it properly, and persist it.

Mode Selection

Signal	Mode	Phases	When
Casual question ("뭘 배웠지?", "learnings")	Quick	1a → 3 → 6	Simple summary, end-of-session
Maintenance request ("메모리 정리", "cleanup")	Full	1~6 all	Memory audit, bulk operations
Track argument ("/retrospect OR")	Full	1~6 all	Deep track analysis
During retrospective (③ step)	Full	1~6 all	Post-audit protocol

Mode is selected automatically — do not ask the user.

Execution Context

Context	Detection	Approval Behavior
Interactive	Main session, user is responding	Present candidates → wait for approval → execute
Headless	Subagent, no human to respond	Auto-approve threshold-met candidates → defer ambiguous → execute → report

In headless mode, NEVER ask questions or wait for input. Any prompt will block indefinitely. Auto-approve what meets threshold rules, defer the rest as ⏭️ deferred to orchestrator.

Setup

Read config: ${CLAUDE_PLUGIN_ROOT}/config.json → plugin.locale

Locate memory: use Glob to find **/memory/MEMORY.md. Read it to understand existing memories (avoid duplicates, check line count < 200).

Phases

Phase 1: Gather Sources

Read references/gathering.md for detailed source collection methods.

Quick mode: conversation + git log only (skip audit_history). Full mode: conversation + audit_history + git log.

Phase 2: Deduplicate (Full mode only)

Read existing memory files → build topic index → mark candidates as "new" or "update".

Phase 3: Generate Candidates

Read references/candidates.md for memory format template, categories, and threshold rules.

4 categories: feedback (corrections + confirmations), project (work context), user (role/preferences), reference (external pointers).

Phase 4-6: Present → Execute → Verify

Read references/execution.md for presentation format, approval flow, and integrity checks.

Interactive: present table → wait for approval
Headless: auto-approve → execute → report with deferred items

Integration with Retrospective

When invoked during post-audit retrospective (③ Memory cleanup step):

This skill handles memory extraction and cleanup
Other retro steps (① what went well, ② problems, ④ feedback, ⑤ Act) continue independently
Learnings needing code changes → work-catalog items (not memories)

Memory Authority

Only this skill and agents may read/write memories. Other skills (verify, merge, planner) must not access the memory directory directly. This prevents memory pollution from non-learning contexts.

Actor	Memory Access
`consensus-loop:retrospect`	Read + Write (primary owner)
Agents (orchestrator, implementer)	Read only (via auto-memory system)
Other skills (verify, merge, planner)	No access — request via orchestrator if needed

Rules

Data before candidates — gather all sources before generating
Deduplicate before presenting — check existing memories first (Full mode)
Every candidate cites a source — no unsupported claims
User approves everything (interactive) / auto-approve with thresholds (headless)
Why + How to apply required for feedback/project types
Prefer update over create — one good memory > two overlapping
Show what you skipped — user may disagree with threshold
Verify after execution — MEMORY.md integrity check (Full mode)
Never block in headless mode — no questions, no prompts, no "should I?"
Memory authority — only this skill writes memories; other skills must not touch memory files

consensus-loop:retrospect