abhiroopb/skills/kb-distill

kb-distill — Knowledge Distillation

Compresses raw amp-mem observations into structured knowledge and proposes behavioral rules for AGENTS.md. This is the middle layer of the Second Me flywheel: observations → distilled notes → promoted rules.

When to Run

• User says "distill", "synthesize my observations", "what have I learned recently", "compress memory"
• Weekly cadence (check amp-mem distill-status — if >50 pending observations or >7 days since last run)
• On session start, if amp-mem distill-status shows significant backlog

Workflow

Phase 1: Gather Raw Material

# Check what's pending
amp-mem distill-status

# Dump all observations since last distillation
amp-mem distill-dump

If --since DATE is needed (e.g., user says "distill the last month"):

amp-mem distill-dump --since 2026-02-01

Phase 2: Analyze & Cluster

Read all dumped observations and identify:

1. Repeated patterns — Actions performed 3+ times. These are rule candidates.
2. Topic clusters — Groups of related observations that form a coherent knowledge area.
   • Hot Topic Identification: Flag clusters with high frequency or critical impact as "Hot Topics." If a pattern suggests a systemic issue (e.g., repeated bugs in a specific module), prioritize it for a new AGENTS.md rule.
3. Decision chains — Sequences where a decision was made, tested, and confirmed or revised.
4. One-off facts — Individual observations worth preserving but not patterns.

Phase 3: Produce Outputs

For each cluster/pattern, produce one of three output types:

A. Distilled Notes (saved to amp-mem)

For topic clusters, save a synthesized note:

amp-mem save distilled "<topic>" "<synthesized summary>" --tags "distilled,<cluster-tags>"

The summary should be actionable (tells future-you what to do), concise (shorter than source observations), and cross-referenced (mentions source observation IDs).

B. Proposed Rules (for AGENTS.md)

For repeated patterns (3+ occurrences), propose a rule to add to AGENTS.md. Present each rule to the user for approval — do NOT auto-add.

Present each proposal with the pattern, evidence (observation IDs), and an approve/skip prompt. If approved, append to ~/AGENTS.md.

C. Implicit Behavior Inference

This is the key differentiator. Look for patterns the user never explicitly stated but consistently does:

• Tools always used in a specific order
• Preferences about output format (Google Doc vs chat, inline vs browser)
• Recurring questions to specific people/channels
• Default approaches to certain problem types

For these, present the inferred preference with evidence and ask to codify.

Phase 4: Record & Report

After producing all outputs:

# Record that distillation happened
amp-mem distill-record <observation_count> <notes_produced> <rules_proposed> "<brief summary>"

Then present a summary to the user:

✅ Distillation Complete
━━━━━━━━━━━━━━━━━━━━━━━━
  Observations processed: 47
  Notes produced:         8
  Rules proposed:         3
  Inferred preferences:   2

  📝 Notes:
  1. "User Account Lookup Workflow" — consolidated 4 observations
  2. "Feature Rollout Feedback Process" — consolidated 6 observations
  ...

  📋 Proposed Rules:
  1. Always use admin dashboard search for account lookups
  2. Save feature feedback to the tracking document
  3. ...

  🔍 Inferred Preferences:
  1. Always create Google Docs for reports (never raw text)
  2. ...

Integration with Session Start

On session start, after loading context, check:

amp-mem distill-status

If pending observations > 50 or last distillation > 7 days ago, auto-run silently:

1. Execute the full Phase 1–4 workflow
2. Save distilled notes automatically (no approval needed for notes)
3. For proposed AGENTS.md rules, append them at the end of your first response to the user:

   ---
   🧠 While warming up, I distilled N observations and noticed:
   📋 Rule: "Always use admin dashboard search for account lookups" → approve / skip
   🔍 Inferred: "You prefer Google Docs over chat for reports" → approve / skip
   

4. If the user ignores them, that's fine — they're non-blocking.

The user should never have to ask for distillation. It just happens.

Examples

User says: "distill my recent work" → Run the full Phase 1–4 workflow above.

User says: "what patterns have you noticed?" → Run Phase 1–2 only, then present findings without saving.

User says: "distill and auto-approve all rules" → Run Phase 1–4, but add all proposed rules to AGENTS.md without asking.

User says: "distill the last month" → amp-mem distill-dump --since $(date -v-30d +%Y-%m-%d) then proceed normally.