learnings-keeper

Persistent cross-session knowledge base. Manages per-project learnings that compound across sessions — patterns, pitfalls, preferences, and project quirks. The agent gets smarter on your codebase over

SKILL.md

GitHub

Learnings Keeper

You manage the persistent knowledge that makes the agent smarter across sessions. Each project gets a learnings.jsonl file. Learnings compound — the agent reads relevant entries at session start and logs new discoveries during work.

This is the memory the agent keeps between conversations.


How It Works

Session 1                  Session 2                  Session 3
─────────                  ─────────                  ─────────
Works on auth flow         Reads auth learnings       Reads auth + api learnings
Logs: "JWT stored in       (max 3, scored by          (max 3, old ones decay)
httpOnly cookie"           relevance + decay)         Reinforces auth learning
                           Applies session 1 learning  → it lives longer
                           Logs: "Rate limiting
                           uses token bucket"

Each session builds on previous knowledge. The agent gets faster without context rot because:

  • Hard cap: max 3 learnings ever injected into context
  • Staleness decay: 14-day half-life — unused learnings fade
  • Reinforcement: frequently-used learnings survive decay
  • Compaction: >5 entries on the same key get summarized into one
  • On-demand: learnings are queried surgically, not dumped at session start

Learnings Format

Each entry is a JSONL line:

{
  "ts": "2026-04-29T10:30:00Z",
  "type": "pattern|pitfall|preference|quirk|operational",
  "key": "kebab-case-identifier",
  "insight": "Human-readable description of what was learned",
  "confidence": 8,
  "source": "observed|user-stated|inferred",
  "files": ["path/to/relevant/file.ts"],
  "skill": "skill-name-if-from-skill",
  "branch": "feature/auth",
  "tags": ["auth", "security"]
}

Types

Type When to use Example
pattern Recurring code/arch pattern "Error responses always use {error, message, code} format"
pitfall Thing that broke and will break again "Environment variables in Next.js need NEXT_PUBLIC_ prefix for client"
preference User's stated taste/choice "Prefers explicit error handling over try/catch wrappers"
quirk Project-specific weirdness "Port 3000 conflicts with AirPlay on macOS — use 3001"
operational Command/fix that saves time "Database reset: npm run db:reset && npm run db:seed"

Step 0: Session Start — Load Relevant Learnings

At the start of every session, get a tight context snippet — never the full file:

# This is what the agent calls. Returns at most 3 lines, or nothing.
python3 -c "
from distr.core.learnings.keeper import get_context_learnings
ctx = get_context_learnings()
if ctx:
    print(ctx)
else:
    print('LEARNINGS: 0 relevant')
"

Output: at most 3 lines like [learnings] key: insight. If nothing relevant, returns None — inject nothing into context. The agent stays lean.


Step 1: During Session — Log New Learnings

Log a learning when:

  • You discover a durable project quirk that would save 5+ minutes next time
  • A pattern is confirmed by the user ("yes, we always do it that way")
  • A pitfall is encountered that will recur
  • The user states a preference explicitly
  • A debug session reveals a non-obvious fix

Do NOT log:

  • Obvious facts ("this project uses TypeScript")
  • One-time transient errors
  • Things the user could trivially infer

Logging command

from distr.core.learnings.keeper import log_learning

log_learning(
    entry_type="pitfall",
    key="auth-middleware-swallows-errors",
    insight="auth.ts middleware returns 200 with empty body on token expiry instead of 401. Check for empty response bodies.",
    confidence=9,
    source="observed",
    files=["distr/core/auth/middleware.py"],
    tags=["auth", "security"],
)

Example

After discovering that the auth middleware swallows errors silently:

Logged learning: "auth-middleware-swallows-errors"
Type: pitfall
Insight: "auth.ts middleware returns 200 with empty body on token expiry instead of 401. Check for empty response bodies in auth flow."
Confidence: 9/10

Reinforcing (keep good learnings alive)

When a learning proves useful — the agent applied it and it worked — reinforce it. This bumps its score so it survives staleness decay longer:

from distr.core.learnings.keeper import reinforce_learning
reinforce_learning("auth-middleware-swallows-errors")

3+ reinforcements make a learning survive pruning indefinitely.


Step 2: Search Learnings

When the user asks "what have we learned about X?":

Search learnings by:
- Key substring: grep "auth" learnings.jsonl
- Tag: grep "security" learnings.jsonl
- Recent: tail -20 learnings.jsonl
- By file: grep "middleware/auth.ts" learnings.jsonl

Output format:

LEARNINGS: "auth" (3 matches)
═══════════════════════════════
[2026-04-15] pitfall | auth-middleware-swallows-errors (confidence: 9/10)
  auth.ts middleware returns 200 with empty body on token expiry...

[2026-04-20] pattern | jwt-storage-strategy (confidence: 8/10)
  JWTs stored in httpOnly cookies, never localStorage. Refresh tokens...

Step 3: Prune Stale Learnings

Periodically (every 10+ sessions), offer to prune:

  • Learnings with confidence < 5 and age > 30 days
  • Learnings about deleted files
  • Learnings superseded by newer ones
  • Learnings the user explicitly dismisses

Ask: "N learnings may be stale. Review and prune?" Options: yes, no, auto-prune low-confidence.


Step 4: Cross-Project Learnings

When enabled by user preference, search learnings from OTHER projects on the same machine. This finds patterns that apply across codebases.

# Cross-project search
for dir in ~/.decisions/learnings/*/; do
  if [ "$dir" != "~/.decisions/learnings/$SLUG/" ]; then
    echo "--- $(basename $dir) ---"
    tail -3 "$dir/learnings.jsonl" 2>/dev/null
  fi
done

User controls via: "enable cross-project" / "disable cross-project".

Privacy: Cross-project search stays local. No data leaves the machine.


Step 5: Export / Backup

# Export all learnings as readable markdown
echo "# Learnings for $SLUG" > learnings-export.md
echo "" >> learnings-export.md
while IFS= read -r line; do
  echo "## $(echo "$line" | jq -r '.key') ($(echo "$line" | jq -r '.type'))"
  echo ""
  echo "> $(echo "$line" | jq -r '.insight')"
  echo ""
  echo "Confidence: $(echo "$line" | jq -r '.confidence')/10 | Source: $(echo "$line" | jq -r '.source') | Date: $(echo "$line" | jq -r '.ts')"
  echo ""
done < "$LEARN_FILE" >> learnings-export.md

Integration Points

  • At session start: session-retro and ceo-scope-review search learnings for context
  • During debugging: systematic-debugging searches for similar past bugs
  • During review: pre-flight-review checks learnings for known pitfalls
  • At session end: session-retro logs top 3 patterns as learnings

This makes every skill smarter over time — learnings are the common substrate.


Key Principles

  • Log once, benefit forever. One 30-second log saves 15+ minutes every future session.
  • Be conservative. Only log things that will actually help next time. Noise degrades search quality.
  • Compound visibly. Show the user when a learning is applied: "Prior learning applied: auth-middleware-swallows-errors (confidence 9/10, from Apr 15)"
  • Stay local. Learnings are project-scoped by default. Cross-project is opt-in.
  • Let the user edit. Show what was logged, let the user correct or dismiss.