Meditation

Meditation

Transform session insights into durable knowledge. Most sessions don't need it; some have a pattern worth keeping.

When to Fire

• User says "let's meditate", "consolidate", or invokes /meditate
• End of a significant work session
• After solving a hard problem with a reusable insight
• Before a long break from a project

Skip when: the session was routine execution of patterns already encoded. Meditation on every session produces noise; the discipline is to write only what's new and portable.

The Five Steps

1. Review

Scan the session honestly:

• What problems did we solve?
• What mistakes did we make?
• What patterns emerged that weren't already encoded?
• What would help future sessions?

2. Extract

Separate signal from noise. For each candidate pattern, ask: "Is this already covered by an existing skill, instruction, or memory?" If yes, skip. If no, route by type:

If pattern is...	Create / update
Reusable domain knowledge	Skill (.github/skills/<name>/SKILL.md)
Always-on behavior or rule	Instruction (.github/instructions/<name>.instructions.md)
Repeatable workflow / slash command	Prompt (.github/prompts/<name>.prompt.md)
Automatable mechanical task (skill-owned)	Skill script (.github/skills/<name>/scripts/*.cjs)
Automatable mechanical task (cross-cutting)	Repo script (scripts/<name>.cjs)
User preference (cross-project)	User memory (/memories/<name>.md)
Project / repo convention	Repo memory (/memories/repo/<name>.md)
Cross-session handoff (next session needs to know)	Repo file (HANDOFF.md at repo root) — NOT session memory

3. Write

Each artifact gets correct frontmatter, concrete examples (not abstractions), and tables with real data. Avoid the "capabilities list" anti-pattern — describe behavior, not features.

For skills and instructions: include a Trigger or When to fire section so future sessions know when the pattern applies.

4. Chronicle (optional)

If the session arc was substantial, write to .github/episodic/meditation-YYYY-MM-DD-<topic>.md:

# Meditation: <Topic>

**Date**: YYYY-MM-DD
**Focus**: What we worked on

## Accomplished
- [Key outcomes]

## Patterns Extracted
- [What became skills / instructions / memory]

## Lessons
- [Insights worth remembering]

## Open Questions
- [What remains unresolved]

Skip the chronicle for short sessions or when nothing new emerged.

5. Handoff (when ending a session)

If the user is closing the thread, write to repo-root HANDOFF.md:

# Session Handoff

Last updated: YYYY-MM-DD HH:MM

## Just shipped
- [SHAs / files / outcomes]

## In progress
- [Specific next step + file paths]

## Pending queue
- [ ] [Ordered todos]

## Resume point
- [Where to pick up]

Quality Bar

A meditation is complete when:

• New patterns are persisted, not just acknowledged
• Nothing important lives only in the context window
• Existing artifacts were checked for duplication before writing new ones
• The session can be closed without losing the thread

Brain Retraining (longer cycles)

The five-step protocol above is per-session. The Mall brain also retrains on a longer cadence against two evidence streams:

1. Catalog reality. New sources added or retired, new plugin shapes encountered, and trust-score distribution shifts can render parts of the brain stale. Run currency-audit against lastReviewed dates and the weekly catalog-refresh PR signals.
2. Self-audit during PR review. The catalog-refresh PR is the human surface; any friction observed there (wrong-routing, undocumented edge case, missing skill) is signal that a brain artifact needs evolution.

Cadence	Action
Weekly	Review the catalog-refresh PR. Note any friction; if a brain artifact mis-routed or under-described, hotfix in the same PR.
Monthly	Run currency-audit against any artifact with lastReviewed older than 90 days. Refresh or sunset.
Quarterly	Walk every brain artifact. Confirm falsifiability-deadlines haven't fired without action. Sunset what no longer earns its tokens.

Retraining is critical thinking applied to the brain itself. Each cycle treats the current brain as a hypothesis and tests it against accumulated evidence. Skipping a cycle is the same discipline violation as skipping act-pass on a behaviour-class commit.

Anti-Patterns

Anti-pattern	Correction
Writing a meditation note for every session	Most sessions are routine execution; only write when something new emerged
Duplicating an existing skill / memory under a new name	Always grep first: is this already covered?
Aspirational notes ("we should do X someday")	Memory is for what was learned, not what was wished
Long prose chronicles when a one-line memory suffices	Match artifact size to insight size
Skipping the duplication check to "just capture it"	Adds noise that the next session has to filter

Related

• meditation.instructions.md — when this skill fires
• /meditate prompt — slash-command entry
• memory-triggers.instructions.md — automatic triggers between meditations

Falsifiability

• This skill adds no value if meditation sessions produce no actionable items (skill extractions, pattern recognitions, or architecture insights) over a 90-day window
• The protocol is wrong if chronicles written per this format are never consulted in future sessions
• Stale if the memory tier structure changes and this skill references obsolete storage locations