Bumblebiber/hmem-wipe

Wipe — Prepare & Clear Context

Follow these steps in order.

Step 1: Optionally save high-value knowledge

Check checkpointMode in hmem.config.json to decide what to do:

checkpointMode: "auto"

The Haiku subagent already extracts L/D/E entries every 20 exchanges automatically. Skip manual writes unless you have specific high-value knowledge that:

• Was discovered in the last few exchanges (too recent for the last auto-checkpoint)
• Is critical enough that losing it would cost significant rework
• Is NOT already covered by a recent auto-checkpoint

If nothing qualifies, proceed directly to Step 2.

checkpointMode: "remind"

Manually save unsaved insights from this project context:

• New lessons learned: write_memory(prefix="L", ...)
• Project progress: append_memory(id="P00XX.7", ...) (Protocol node)
• Decisions made: write_memory(prefix="D", ...)
• Errors encountered: write_memory(prefix="E", ...)

Skip if the last checkpoint was fewer than 5 messages ago.

Why gate on checkpointMode?

Redundant writes waste tokens and create duplicates that clutter memory. Auto-checkpoints already call read_memory to deduplicate before writing — manual writes during wipe bypass that check and risk creating noise.

Step 2: Update Next Steps

Before clearing context, ensure the active project's "Next Steps" section is up to date. This is critical for session handoff — after /clear, the next session (or restored context) needs to know what to work on next.

1. Find the "Next Steps" section: read_memory(id="P00XX") at depth 2 to list L2 children, then identify the "Next Steps" node by title (seq may vary per project).
2. Review current content: read_memory(id="P00XX.N") where N is the Next Steps seq.
3. Update with current priorities: write_memory(id="P00XX.N", content="...") with:
   • What was being worked on
   • What's done vs. still open
   • Immediate next actions for the next session
   • Any blockers or decisions pending
4. Mark completed steps as irrelevant: update_memory(id="P00XX.N.M", irrelevant=true) for each L3 child under Next Steps that has been fully completed. This keeps the section clean for the next session — irrelevant nodes are hidden from load_project output.

Skip if Next Steps is already current (updated within the last few exchanges).

Step 3: Tell the user to /clear

O-entries are auto-logged by the Stop hook — every exchange is already saved to the active project's O-entry. No need to manually create O-entries or call flush_context for conversation history.

Reply with exactly:

Context ready for clear. Type /clear — the SessionStart hook will automatically restore your project context.

Do NOT attempt to run /clear yourself — it is a built-in CLI command only the user can execute.

What happens after /clear

Context is restored automatically by the SessionStart[clear] hook — no agent action needed after /clear. Do NOT call load_project or read_memory during or after this skill; the next session's first UserPromptSubmit hook will trigger the normal hmem-read flow with a verified active project ID.

The hook:

1. Resets the MCP session cache
2. Injects recent conversation exchanges from the project's O-entry transcript
3. Injects the active project briefing (overview expanded)
4. Injects recent O-entry titles + rules

Why this flow works

• O-entries are covered. The Stop hook logs every exchange to the active project's O-entry. Wipe does not need to handle conversation history.
• Checkpoints are covered (auto mode). The Haiku subagent extracts knowledge every 20 exchanges. Wipe only needs to catch the tail end, if anything.
• Context restoration is covered. The SessionStart[clear] hook handles re-injection automatically. The agent just needs the user to type /clear.