recall-learnings

name	recall-learnings
description	Surface prior learnings relevant to the current task. Use before starting any non-trivial work — new experiment, new feature, new submission strategy, debugging a pattern — to check if we've been here before. Also use when the user says "what have we learned about X", "any past notes on Y", "recall", or "check learnings".

Recalling Prior Learnings

Surfaces the most relevant learnings from .learnings/ for the task at hand. This is the compound-interest mechanism — every session should call it before committing to an approach.

When to invoke proactively

Before:

• Picking a modeling approach ("LightGBM vs CatBoost for tabular").
• Designing a CV split strategy.
• Writing submission glue code for a new competition.
• Starting an agent's main decision loop.
• Deciding a hyperparameter range.
• Writing code that touches the Kaggle CLI / MCP.

Basically: any time you would otherwise guess, recall first.

Steps

1. Build a query. Extract 3-6 keywords from the task at hand. Examples:
   • "class imbalance" + "lightgbm" + "tabular"
   • "agent timeout" + "simulation" + "move budget"
   • "submission format" + "csv" + "leaderboard regression"
2. Scope filter. If inside competitions/<slug>/, include both scope: global and scope: <slug>. Otherwise global only.
3. Search:

   # tag-first (precise)
   grep -l -E "tags:.*\b(<tag1>|<tag2>)\b" .learnings/L-*.md
   # keyword fallback
   grep -l -i -E "<keyword1>|<keyword2>" .learnings/L-*.md
   

4. Rank by: recurring=true first, then cost: high, then severity: scar, then recency. Cap at 5.
5. Read the top matches. Extract the Rule lines only.
6. Present to the user as a compact numbered list: [ID] Rule — (scope, cost).
7. Apply. When proceeding, explicitly cite the learnings you are honoring: "Applying L-… by using stratified k-fold instead of random split."

Example output

Relevant prior learnings:
  1. L-2026-03-02-001 — Cast submission ID columns with .astype(int) before writing (global, high-cost)
  2. L-2026-03-15-003 — Stratified-by-target beats random split on tabular competitions (global, medium)
  3. L-2026-03-20-002 — Leaderboard-driven overfit: trust CV when delta > 0.01 (global, high-cost)
Applying: casting IDs, stratified split.

Anti-patterns

• Don't dump every matching learning verbatim — pick the top 3-5 and show only their Rule lines.
• Don't invent learnings that "feel right." If nothing matches, say so; that's honest and opens a slot for a new capture later.
• Don't recall, list, and then ignore — citing a learning in the plan is what enforces discipline.