Load Learnings

Load Learnings (v2 — Memory Graph)

Hard cap: 15 memories per session. No exceptions.

Protocol

Step 1 — Load scored memories

Use the memory_load MCP tool:

memory_load({
  project: "{current project name}",
  activeSkills: ["next-best-practices", "systematic-debugging", ...],
  limit: 15,
  repo: "skillbrain"
})

The scoring algorithm calculates:

score = (confidence × 2)
      + (3 if scope=global OR project matches)
      + (2 if validated within 5 sessions)
      + (2 if memory belongs to active skill)
      + (importance × 0.5)

Step 2 — Supplement with semantic search (optional)

If the current task has a clear focus, search for additional relevant memories:

memory_search({
  query: "{task description in natural language}",
  limit: 5,
  repo: "skillbrain"
})

Merge with scored results, dedup by ID, cap at 15.

Step 3 — Load into context

For each selected memory, use these fields:

• context — when/where this applies
• problem — what went wrong
• solution — actionable fix
• confidence (trust signal: 1-3 = tentative, 4-7 = reliable, 8-10 = established)
• type — memory type (Pattern, BugFix, etc.)

Step 4 — Present summary

The memory_load tool returns a formatted summary:

📚 Loaded {N} memories
   Pattern: {N}
   BugFix: {N}
   AntiPattern: {N}
   ...

Add trust indicators:

   ⚠️ Tentative (confidence ≤ 2): {N} — treat as suggestions, not rules
   ✅ Established (confidence ≥ 7): {N} — highly trusted

When to reload mid-session

• Session shifts significantly (debugging → UI work) → reload with new focus
• Project switches → full reload with new project context
• Do NOT reload for every small task change — only major context shifts

Fallback

If Memory Graph is unavailable (database error, etc.), fall back to grep-based search:

grep -r "confidence: [4-9]\|confidence: 10" \
  ".agents/skills/*/learnings.md" ".opencode/skill/*/learnings.md" \
  -A 5 -B 10 2>/dev/null | head -30