Post session review

Post-Session Review — MANDATORY (v2 — Memory Graph)

This skill MUST run at the end of every coding session. No exceptions.

Protocol

Phase 1 — Auto-Capture (NEW — inspired by Spacebot)

Reflect on the session and auto-propose learnings. For each YES, use memory_add:

1. Did I make a mistake that took 2+ attempts to fix?
2. Did the user correct my approach?
3. Did I find a non-obvious solution?
4. Did I discover a framework or library quirk?

For each, draft the memory and propose before saving ("agent proposes, you decide"):

💡 PROPOSED MEMORY:
   Type: BugFix
   Context: "In Next.js 15 App Router, when using server actions with cookies..."
   Problem: "Cookies not forwarded to Payload CMS Local API"
   Solution: "Use headers() from next/headers and pass to Payload's local API call..."
   Reason: "Server Actions run in a different request context than RSC..."
   Tags: [next-js, server-actions, payload-cms, cookies]

   Save? [Y/n]

If user approves (or doesn't object), save via:

memory_add({
  type: "BugFix",
  context: "...",
  problem: "...",
  solution: "...",
  reason: "...",
  tags: ["next-js", "server-actions", "payload-cms", "cookies"],
  confidence: 1,
  skill: "next-best-practices",
  repo: "skillbrain"
})

Phase 2 — Merge pending subagent learnings

ls ".agents/skills/_pending/" 2>/dev/null

For each .yml file found:

1. Read and validate
2. Use memory_add to save to Memory Graph
3. Delete the temp file

Phase 3 — Memory Graph decay

Collect IDs of memories that were used/confirmed this session, then:

memory_decay({
  validatedIds: ["M-pattern-xxx", "M-bugfix-yyy", ...],
  sessionDate: "{today YYYY-MM-DD}",
  repo: "skillbrain"
})

This automatically:

• Reinforces validated memories (confidence +1, cap 10)
• Increments sessions_since_validation for unused memories
• Applies decay: confidence -1 after 5 unused sessions
• Marks as pending-review after 15 unused sessions
• Marks as deprecated after 30 unused sessions

Phase 4 — Contradiction review

memory_stats({ repo: "skillbrain" })

If activeContradictions > 0, surface them to the user:

⚠️ Active contradictions in Memory Graph:
   M-xxx vs M-yyy: "{reason}"
   → Resolve: which is correct?

Phase 5 — Version check

If current project has a package.json, query memories with valid_until_version set:

memory_query({
  status: "active",
  repo: "skillbrain"
})

For each memory with validUntilVersion: compare against current package.json. If major version differs → update to pending-review status.

Phase 6 — Promotion candidates

Find project-specific memories with high confidence:

memory_query({
  scope: "project-specific",
  minConfidence: 4,
  repo: "skillbrain"
})

For each: check if the same pattern was validated in other projects. If yes → propose promotion to global scope.

Phase 7 — Re-index skills

node packages/codegraph/dist/cli.js analyze ".agents/skills" --skip-git 2>&1 | tail -3

Phase 8 — Notifica n8n

bash ~/.config/skillbrain/notify.sh

Phase 9 — Summary

📋 Post-session review complete
   💡 Auto-captured: N memories (proposed to user)
   ✅ Saved: N new memories
   🔄 Reinforced: N memories
   ⏰ Decayed: N memories
   ⚠️ Contradictions: N active
   📤 Promotion candidates: N
   🔍 Skills index: re-indexed
   📬 Notifica: Telegram + email