learn-from-session

name	learn-from-session
description	Use when ending a session, when the user asks to save lessons, or when you notice a correction or preference worth remembering. Captures non-obvious lessons to AGENTS.md or skill files so the agent gets smarter over time. Do NOT use for trivial facts derivable from reading code.

Core Workflow

1. Scan the session — re-read the conversation. Identify non-obvious lessons:

   • Corrections — agent did X, user said "no, do Y instead"
   • Validated approaches — user confirmed a non-obvious choice without pushback
   • Revealed preferences — how the user likes to work, what they care about, what annoys them
   • Project facts — decisions, constraints, context not in code or git history

2. Categorize each lesson — use the memory conventions tree:

   Is this a universal rule for a language, framework, or tool?
   ├── Yes → the relevant global skill in ~/.agents/skills/
   └── No  → Is it about user preferences or codebase conventions?
            ├── Yes → {repo}/AGENTS.md
            └── No  → Skip
   

3. Present to user — use the question tool. For each lesson, show a short description and proposed destination. Let the user confirm or redirect before writing.

4. Save — write the confirmed lessons to the appropriate files. Append to the relevant section.

5. Report — briefly state what was found (or "nothing worth saving") and where each lesson was saved.

Common Rationalizations

Rationalization	Reality
"This is a standard best practice, everyone knows this"	If the agent violated it and got corrected, it's not standard for this agent. Save it.
"I'll remember this for next time"	You won't. Write it down now.
"The user didn't say this explicitly, I inferred it"	Don't save inferred preferences — only explicit corrections and confirmations.
"This is already in my training data"	Then why did you get it wrong? The user's correction takes priority over training data.

Red Flags

• You're saving things derivable from reading the current code
• You're saving a one-off task detail with no future relevance
• You're saving an inference, not an explicit correction
• You're saving without user confirmation
• You're saving to the wrong destination (user preference vs. global skill)

Verification

• Every saved lesson was explicitly confirmed by the user
• Saved to the correct destination per the memory conventions tree
• Nothing was saved that's derivable from reading the code
• User received a summary of what was saved and where

Constraints

MUST DO

• Ask the user before saving — never save without confirmation.
• Use the memory conventions tree to determine destination.
• Capture explicit corrections and confirmations only — no inferred preferences.
• Save one lesson at a time — each gets its own confirmation.
• Report what was saved and where.

MUST NOT DO

• Save things derivable from reading the current code.
• Save inferences as facts.
• Save without asking.
• Batch multiple lessons into one confirmation.
• Save to global skills for project-specific preferences.