learn-recall

Compound Learnings — Recall

Consult the agent's own accumulated work-lessons (corrections, playbooks, insights, patterns) to answer "have we learned this before?" and to inform how you approach the current task. These are the agent's lessons — not source wiki pages (wiki-query handles those) and not user preferences (those live in MEMORY).

This skill is index-first and token-frugal: read the compact index, then fetch at most a handful of bodies, and only the ones that actually sharpen the answer. Never load the store wholesale.

When to Use This Skill

• User asks what was learned/decided before, or references past mistakes, corrections, playbooks, or "lessons learned."
• User asks "have we hit this before?", "did I already solve this?", "what's our approach to X?"
• You are about to start a task and want to check for a relevant past correction or blessed playbook before acting.

When NOT to Use This Skill

• Domain questions about ingested sources → use wiki-query (the wiki).
• User preferences/behaviors → those are MEMORY, not learnings.
• Capturing a new lesson → use /learn-capture. This skill is read-only.
• General coding questions with no relevant captured learning.

The Two Stores

Retrieval merges both tiers; the project store shadows global on id collision.

Scope	Root	Notes
project	<repo>/.compound/	primary, committed, team-shared
global	$COMPOUND_KNOWLEDGE_HOME or ~/.claude/compound-knowledge/	opt-in, curated, cross-project

Each store: a compact index.md plus per-type bodies under insights/ playbooks/ corrections/ patterns/. The index line schema is:

- [CODE] {id} | {tags} | {headline} | {confidence} | {date}

Type codes → body directory: C → corrections/, P → playbooks/, I → insights/, Pa → patterns/. The body file is <root>/<dir>/<id>.md.

Recall Procedure

1. Locate the indexes. Read <repo>/.compound/index.md and the global index.md (under $COMPOUND_KNOWLEDGE_HOME or ~/.claude/compound-knowledge/). If neither exists, say no learnings have been captured yet and stop — there is nothing to recall.

2. Scan the index lines (cheap — they are headline-only). Match the user's question against the tags column and headlines. If a store is large, use grep -i "<term>" <root>/index.md to find candidate lines rather than reading the whole file. Corrections rank first (mistakes not to repeat), then playbooks, then insights/patterns.

3. Merge tiers, project over global on duplicate id. Keep the most relevant candidates (typically ≤5).

4. Fetch bodies on demand — at most 5, only when needed. For a candidate that directly answers the question, read its body at <root>/<type-dir>/<id>.md (or Glob for **/<id>.md if unsure of the type). Skip bodies whose headline already answers the question — many recalls need no body read at all.

5. Synthesize and cite ids and scope. Lead with corrections. For each lesson used, cite its id and scope, e.g.:

   • "A past correction (kw-2026-06-08-retry-jitter, project) says: add jitter…"
   • "Per the blessed playbook kw-…-deploy (global): …" Distinguish a captured learning from general reasoning when you add your own.

Guardrails

• Index-first, never load the store wholesale. Read indexes, then bodies on demand.
• At most 5 bodies per recall unless the user explicitly asks for a full review.
• Read-only. This skill never writes, edits, or archives. To capture or update a lesson, route to /learn-capture.
• Cite ids + scope for every lesson you rely on.
• Corrections first — surface contradicting/cautionary lessons before reinforcing ones.