Memory Sweep

Memory Sweep - find memory-only rules that should be in copilot-instructions.md

When to use this skill

Trigger this skill whenever the user asks any of:

• "audit my memories"
• "sweep my memories"
• "any memory-only rules?"
• "review my Copilot memories"
• "what memories should be promoted to instructions"
• "is anything in memory that should be in the file?"

Also proactively offer to run this skill if you notice during a session that a user-stated rule lives only in memory and not in the instructions file (e.g., when a recent multi-model PR review flags a "smuggled rule" finding like the gratitude-first case).

Why this matters

Memories are great for capturing preferences mid-conversation, but they have downsides:

• They're personal to one user account and don't transfer to teammates.
• They can be down-voted into oblivion accidentally.
• They're invisible during code review or onboarding.
• They can drift out of sync with the canonical instructions file.

Rules that the user wants to apply consistently belong in copilot-instructions.md, where they're visible, source-controlled, and survive any memory churn. The memory remains useful as a quick reference but the file is authoritative.

How to run the sweep

Follow these steps in order:

Step 1: Dump the current memories block

You have the user's stored memories in your prompt context inside a <memories> block. Write that block verbatim to a tempfile in a private per-invocation directory. Memories can contain personal preferences and context you do not want on a predictable shared path.

SWEEP_DIR=$(mktemp -d -t memory-sweep.XXXXXX)
chmod 700 "$SWEEP_DIR"

Then use your file-write tool to create "$SWEEP_DIR/memories.md" containing only the memory entries from the <memories> block - one entry per memory, in this exact markdown format:

**subject heading**
- Fact: <fact text>
- Citations: <citations text>

**next subject**
- Fact: <fact text>
- Citations: <citations text>

Do not include the surrounding instructional prose ("Be sure to consider these stored facts carefully...", "If you come across a memory you can verify...", etc.). Only the memory entries themselves.

Step 2: Run the classifier

python3 ~/.copilot/skills/memory-sweep/sweep.py \
  "$SWEEP_DIR/memories.md" \
  ~/repos/dotfiles/.github/copilot-instructions.md

To focus on only the promotion candidates:

python3 ~/.copilot/skills/memory-sweep/sweep.py \
  "$SWEEP_DIR/memories.md" \
  ~/repos/dotfiles/.github/copilot-instructions.md \
  --only PROMOTE

When the review is finished, clean up the temp directory:

rm -rf "$SWEEP_DIR"

Step 3: Review the output with the user

The output sorts findings by score (lowest first), so the strongest PROMOTE candidates appear at the top. For each candidate:

1. Read the fact text aloud (or summarize it).
2. Confirm with the user whether it's a rule they want documented in the instructions file.
3. If yes, ask which section of the file it belongs under (e.g., Writing Style, Pull Requests, GitHub Actions).
4. If the user says "skip" or "leave in memory only", move on.

Do not silently file a PR for everything flagged as PROMOTE. The classifier is a heuristic and the user gets the final say on what becomes documented policy.

Step 4: Promote agreed-upon rules

For each rule the user approves for promotion:

1. Draft the change to ~/repos/dotfiles/.github/copilot-instructions.md. Match the surrounding section's tone and structure.
2. Self-lint with the validate-style skill before opening a PR.
3. Follow the user's PR workflow from their personal instructions: multi-model review, draft PR, sign-off, Co-authored-by Copilot trailer, full repo PR template.

Verdict guide

PRESENT findings are not always perfect matches. If the user has recently added a rule to the file and the memory was already there, the two should agree. If they don't, that's a different problem (drift) worth surfacing.

Caveats

• This is a keyword and phrase classifier, not a semantic one. False positives and false negatives both happen. Treat output as a triage aid, not as gospel.
• The classifier only checks the instructions file passed as the second argument. Repo-level .github/copilot-instructions.md files are not considered.
• The <memories> block in your prompt context already excludes memories outside the current scope. Whatever you dump is the working set.

Source: zkoppert/dotfiles — distributed by TomeVault.